Upload 78 files

README.md

@@ -1,10 +1,10 @@
 ---
-title: Style Bert VITS2 JVNV
+title: Style-Bert-VITS2 JVNV
 emoji: 😡😊😱😫
 colorFrom: blue
 colorTo: red
 sdk: gradio
-sdk_version: 4.12.0
+sdk_version: 4.16.0
 app_file: app.py
 pinned: false
 license: agpl-3.0

app.py

@@ -1,209 +1,41 @@
 import argparse
 import datetime
+import json
 import os
 import sys
-import warnings
+from typing import Optional
 
 import gradio as gr
-import numpy as np
 import torch
-from gradio.processing_utils import convert_to_16_bit_wav
-
-import utils
-from config import config
-from infer import get_net_g, infer
-from tools.log import logger
+import yaml
+
+from common.constants import (
+    DEFAULT_ASSIST_TEXT_WEIGHT,
+    DEFAULT_LENGTH,
+    DEFAULT_LINE_SPLIT,
+    DEFAULT_NOISE,
+    DEFAULT_NOISEW,
+    DEFAULT_SDP_RATIO,
+    DEFAULT_SPLIT_INTERVAL,
+    DEFAULT_STYLE,
+    DEFAULT_STYLE_WEIGHT,
+    Languages,
+)
+from common.log import logger
+from common.tts_model import ModelHolder
+from infer import InvalidToneError
+from text.japanese import g2kata_tone, kata_tone2phone_tone, text_normalize
 
 is_hf_spaces = os.getenv("SYSTEM") == "spaces"
 limit = 100
 
+# Get path settings
+with open(os.path.join("configs", "paths.yml"), "r", encoding="utf-8") as f:
+    path_config: dict[str, str] = yaml.safe_load(f.read())
+    # dataset_root = path_config["dataset_root"]
+    assets_root = path_config["assets_root"]
 
-class Model:
-    def __init__(self, model_path, config_path, style_vec_path, device):
-        self.model_path = model_path
-        self.config_path = config_path
-        self.device = device
-        self.style_vec_path = style_vec_path
-        self.load()
-
-    def load(self):
-        self.hps = utils.get_hparams_from_file(self.config_path)
-        self.spk2id = self.hps.data.spk2id
-        self.num_styles = self.hps.data.num_styles
-        if hasattr(self.hps.data, "style2id"):
-            self.style2id = self.hps.data.style2id
-        else:
-            self.style2id = {str(i): i for i in range(self.num_styles)}
-
-        self.style_vectors = np.load(self.style_vec_path)
-        self.net_g = None
-
-    def load_net_g(self):
-        self.net_g = get_net_g(
-            model_path=self.model_path,
-            version=self.hps.version,
-            device=self.device,
-            hps=self.hps,
-        )
-
-    def get_style_vector(self, style_id, weight=1.0):
-        mean = self.style_vectors[0]
-        style_vec = self.style_vectors[style_id]
-        style_vec = mean + (style_vec - mean) * weight
-        return style_vec
-
-    def get_style_vector_from_audio(self, audio_path, weight=1.0):
-        from style_gen import extract_style_vector
-
-        xvec = extract_style_vector(audio_path)
-        mean = self.style_vectors[0]
-        xvec = mean + (xvec - mean) * weight
-        return xvec
-
-    def infer(
-        self,
-        text,
-        language="JP",
-        sid=0,
-        reference_audio_path=None,
-        sdp_ratio=0.2,
-        noise=0.6,
-        noisew=0.8,
-        length=1.0,
-        line_split=True,
-        split_interval=0.2,
-        style_text="",
-        style_weight=0.7,
-        use_style_text=False,
-        style="0",
-        emotion_weight=1.0,
-    ):
-        if reference_audio_path == "":
-            reference_audio_path = None
-        if style_text == "" or not use_style_text:
-            style_text = None
-
-        if self.net_g is None:
-            self.load_net_g()
-        if reference_audio_path is None:
-            style_id = self.style2id[style]
-            style_vector = self.get_style_vector(style_id, emotion_weight)
-        else:
-            style_vector = self.get_style_vector_from_audio(
-                reference_audio_path, emotion_weight
-            )
-        if not line_split:
-            with torch.no_grad():
-                audio = infer(
-                    text=text,
-                    sdp_ratio=sdp_ratio,
-                    noise_scale=noise,
-                    noise_scale_w=noisew,
-                    length_scale=length,
-                    sid=sid,
-                    language=language,
-                    hps=self.hps,
-                    net_g=self.net_g,
-                    device=self.device,
-                    style_text=style_text,
-                    style_weight=style_weight,
-                    style_vec=style_vector,
-                )
-        else:
-            texts = text.split("\n")
-            texts = [t for t in texts if t != ""]
-            audios = []
-            with torch.no_grad():
-                for i, t in enumerate(texts):
-                    audios.append(
-                        infer(
-                            text=t,
-                            sdp_ratio=sdp_ratio,
-                            noise_scale=noise,
-                            noise_scale_w=noisew,
-                            length_scale=length,
-                            sid=sid,
-                            language=language,
-                            hps=self.hps,
-                            net_g=self.net_g,
-                            device=self.device,
-                            style_text=style_text,
-                            style_weight=style_weight,
-                            style_vec=style_vector,
-                        )
-                    )
-                    if i != len(texts) - 1:
-                        audios.append(np.zeros(int(44100 * split_interval)))
-                audio = np.concatenate(audios)
-            with warnings.catch_warnings():
-                warnings.simplefilter("ignore")
-                audio = convert_to_16_bit_wav(audio)
-        return (self.hps.data.sampling_rate, audio)
-
-
-class ModelHolder:
-    def __init__(self, root_dir, device):
-        self.root_dir = root_dir
-        self.device = device
-        self.model_files_dict = {}
-        self.current_model = None
-        self.model_names = []
-        self.models = []
-        self.refresh()
-
-    def refresh(self):
-        self.model_files_dict = {}
-        self.model_names = []
-        self.current_model = None
-        model_dirs = [
-            d
-            for d in os.listdir(self.root_dir)
-            if os.path.isdir(os.path.join(self.root_dir, d))
-        ]
-        for model_name in model_dirs:
-            model_dir = os.path.join(self.root_dir, model_name)
-            model_files = [
-                os.path.join(model_dir, f)
-                for f in os.listdir(model_dir)
-                if f.endswith(".pth") or f.endswith(".pt") or f.endswith(".safetensors")
-            ]
-            if len(model_files) == 0:
-                logger.info(
-                    f"No model files found in {self.root_dir}/{model_name}, so skip it"
-                )
-            self.model_files_dict[model_name] = model_files
-            self.model_names.append(model_name)
-
-    def load_model(self, model_name, model_path):
-        if model_name not in self.model_files_dict:
-            raise Exception(f"モデル名{model_name}は存在しません")
-        if model_path not in self.model_files_dict[model_name]:
-            raise Exception(f"pthファイル{model_path}は存在しません")
-        self.current_model = Model(
-            model_path=model_path,
-            config_path=os.path.join(self.root_dir, model_name, "config.json"),
-            style_vec_path=os.path.join(self.root_dir, model_name, "style_vectors.npy"),
-            device=self.device,
-        )
-        styles = list(self.current_model.style2id.keys())
-        return (
-            gr.Dropdown(choices=styles, value=styles[0]),
-            gr.update(interactive=True, value="音声合成"),
-        )
-
-    def update_model_files_dropdown(self, model_name):
-        model_files = self.model_files_dict[model_name]
-        return gr.Dropdown(choices=model_files, value=model_files[0])
-
-    def update_model_names_dropdown(self):
-        self.refresh()
-        initial_model_name = self.model_names[0]
-        initial_model_files = self.model_files_dict[initial_model_name]
-        return (
-            gr.Dropdown(choices=self.model_names, value=initial_model_name),
-            gr.Dropdown(choices=initial_model_files, value=initial_model_files[0]),
-            gr.update(interactive=False),  # For tts_button
-        )
+languages = [l.value for l in Languages]
 
 
 def tts_fn(
@@ -216,105 +48,100 @@ def tts_fn(
     length_scale,
     line_split,
     split_interval,
-    style_text,
+    assist_text,
+    assist_text_weight,
+    use_assist_text,
+    style,
     style_weight,
-    use_style_text,
-    emotion,
-    emotion_weight,
+    kata_tone_json_str,
+    use_tone,
+    speaker,
 ):
-    logger.info(f"Start TTS with {language}:\n{text}")
-    logger.info(f"Model: {model_holder.current_model.model_path}")
-    logger.info(f"SDP: {sdp_ratio}, Noise: {noise_scale}, Noise_W: {noise_scale_w}, Length: {length_scale}")
-    logger.info(f"Style text enabled: {use_style_text}, Style text: {style_text}, Style weight: {style_weight}")
-    logger.info(f"Style: {emotion}, Style weight: {emotion_weight}")
-    
     if is_hf_spaces and len(text) > limit:
-        logger.error(f"文字数が{limit}文字を超えています")
-        # raise Exception(f"文字数が{limit}文字を超えています")
-        return f"文字数が{limit}文字を超えています", (44100, "")
-
+        logger.error(f"Text is too long: {len(text)}")
+        return (
+            f"Error: 文字数が多すぎます（{limit}文字以下にしてください）",
+            None,
+            kata_tone_json_str,
+        )
     assert model_holder.current_model is not None
 
+    wrong_tone_message = ""
+    kata_tone: Optional[list[tuple[str, int]]] = None
+    if use_tone and kata_tone_json_str != "":
+        if language != "JP":
+            logger.warning("Only Japanese is supported for tone generation.")
+            wrong_tone_message = "アクセント指定は現在日本語のみ対応しています。"
+        if line_split:
+            logger.warning("Tone generation is not supported for line split.")
+            wrong_tone_message = (
+                "アクセント指定は改行で分けて生成を使わない場合のみ対応しています。"
+            )
+        try:
+            kata_tone = []
+            json_data = json.loads(kata_tone_json_str)
+            # tupleを使うように変換
+            for kana, tone in json_data:
+                assert isinstance(kana, str) and tone in (0, 1), f"{kana}, {tone}"
+                kata_tone.append((kana, tone))
+        except Exception as e:
+            logger.warning(f"Error occurred when parsing kana_tone_json: {e}")
+            wrong_tone_message = f"アクセント指定が不正です: {e}"
+            kata_tone = None
+
+    # toneは実際に音声合成に代入される際のみnot Noneになる
+    tone: Optional[list[int]] = None
+    if kata_tone is not None:
+        phone_tone = kata_tone2phone_tone(kata_tone)
+        tone = [t for _, t in phone_tone]
+
+    speaker_id = model_holder.current_model.spk2id[speaker]
+
     start_time = datetime.datetime.now()
 
-    sr, audio = model_holder.current_model.infer(
-        text=text,
-        language=language,
-        reference_audio_path=reference_audio_path,
-        sdp_ratio=sdp_ratio,
-        noise=noise_scale,
-        noisew=noise_scale_w,
-        length=length_scale,
-        line_split=line_split,
-        split_interval=split_interval,
-        style_text=style_text,
-        style_weight=style_weight,
-        use_style_text=use_style_text,
-        style=emotion,
-        emotion_weight=emotion_weight,
-    )
+    try:
+        sr, audio = model_holder.current_model.infer(
+            text=text,
+            language=language,
+            reference_audio_path=reference_audio_path,
+            sdp_ratio=sdp_ratio,
+            noise=noise_scale,
+            noisew=noise_scale_w,
+            length=length_scale,
+            line_split=line_split,
+            split_interval=split_interval,
+            assist_text=assist_text,
+            assist_text_weight=assist_text_weight,
+            use_assist_text=use_assist_text,
+            style=style,
+            style_weight=style_weight,
+            given_tone=tone,
+            sid=speaker_id,
+        )
+    except InvalidToneError as e:
+        logger.error(f"Tone error: {e}")
+        return f"Error: アクセント指定が不正です:\n{e}", None, kata_tone_json_str
+    except ValueError as e:
+        logger.error(f"Value error: {e}")
+        return f"Error: {e}", None, kata_tone_json_str
 
     end_time = datetime.datetime.now()
     duration = (end_time - start_time).total_seconds()
-    logger.info(f"End TTS, duration: {duration} seconds")
-    return f"Success, time: {duration} seconds.", (sr, audio)
 
+    if tone is None and language == "JP":
+        # アクセント指定に使えるようにアクセント情報を返す
+        norm_text = text_normalize(text)
+        kata_tone = g2kata_tone(norm_text)
+        kata_tone_json_str = json.dumps(kata_tone, ensure_ascii=False)
+    elif tone is None:
+        kata_tone_json_str = ""
+    message = f"Success, time: {duration} seconds."
+    if wrong_tone_message != "":
+        message = wrong_tone_message + "\n" + message
+    return message, (sr, audio), kata_tone_json_str
 
-initial_text = "こんにちは、初めまして。あなたの名前はなんていうの？"
 
-example_local = [
-    [initial_text, "JP"],
-    [
-        """あなたがそんなこと言うなんて、私はとっても嬉しい。
-あなたがそんなこと言うなんて、私はとっても怒ってる。
-あなたがそんなこと言うなんて、私はとっても驚いてる。
-あなたがそんなこと言うなんて、私はとっても辛い。""",
-        "JP",
-    ],
-    [  # ChatGPTに考えてもらった告白セリフ
-        """私、ずっと前からあなたのことを見てきました。あなたの笑顔、優しさ、強さに、心惹かれていたんです。
-友達として過ごす中で、あなたのことがだんだんと特別な存在になっていくのがわかりました。
-えっと、私、あなたのことが好きです！もしよければ、私と付き合ってくれませんか？""",
-        "JP",
-    ],
-    [  # 夏目漱石『吾輩は猫である』
-        """吾輩は猫である。名前はまだ無い。
-どこで生れたかとんと見当がつかぬ。なんでも薄暗いじめじめした所でニャーニャー泣いていた事だけは記憶している。
-吾輩はここで始めて人間というものを見た。しかもあとで聞くと、それは書生という、人間中で一番獰悪な種族であったそうだ。
-この書生というのは時々我々を捕まえて煮て食うという話である。""",
-        "JP",
-    ],
-    [  # 梶井基次郎『桜の樹の下には』
-        """桜の樹の下には屍体が埋まっている！これは信じていいことなんだよ。
-何故って、桜の花があんなにも見事に咲くなんて信じられないことじゃないか。俺はあの美しさが信じられないので、このにさんにち不安だった。
-しかしいま、やっとわかるときが来た。桜の樹の下には屍体が埋まっている。これは信じていいことだ。""",
-        "JP",
-    ],
-    [  # ChatGPTと考えた、感情を表すセリフ
-        """やったー！テストで満点取れた！私とっても嬉しいな！
-どうして私の意見を無視するの？許せない！ムカつく！あんたなんか死ねばいいのに。
-あはははっ！この漫画めっちゃ笑える、見てよこれ、ふふふ、あはは。
-あなたがいなくなって、私は一人になっちゃって、泣いちゃいそうなほど悲しい。""",
-        "JP",
-    ],
-    [  # 上の丁寧語バージョン
-        """やりました！テストで満点取れましたよ！私とっても嬉しいです！
-どうして私の意見を無視するんですか？許せません！ムカつきます！あんたなんか死んでください。
-あはははっ！この漫画めっちゃ笑えます、見てくださいこれ、ふふふ、あはは。
-あなたがいなくなって、私は一人になっちゃって、泣いちゃいそうなほど悲しいです。""",
-        "JP",
-    ],
-    [  # ChatGPTに考えてもらった音声合成の説明文章
-        """音声合成は、機械学習を活用して、テキストから人の声を再現する技術です。この技術は、言語の構造を解析し、それに基づいて音声を生成します。
-この分野の最新の研究成果を使うと、より自然で表現豊かな音声の生成が可能である。深層学習の応用により、感情やアクセントを含む声質の微妙な変化も再現することが出来る。""",
-        "JP",
-    ],
-    [
-        "Speech synthesis is the artificial production of human speech. A computer system used for this purpose is called a speech synthesizer, and can be implemented in software or hardware products.",
-        "EN",
-    ],
-    ["语音合成是人工制造人类语音。用于此目的的计算机系统称为语音合成器，可以通过软件或硬件产品实现。", "ZH"],
-]
+initial_text = "こんにちは、初めまして。あなたの名前はなんていうの？"
 
 example_hf_spaces = [
     [initial_text, "JP"],
@@ -322,30 +149,30 @@ example_hf_spaces = [
     ["吾輩は猫である。名前はまだ無い。", "JP"],
     ["桜の樹の下には屍体が埋まっている！これは信じていいことなんだよ。", "JP"],
     ["やったー！テストで満点取れたよ！私とっても嬉しいな！", "JP"],
-    ["どうして私の意見を無視するの？許せない！ムカつく！あんたなんか死ねばいいのに。", "JP"],
+    [
+        "どうして私の意見を無視するの？許せない！ムカつく！あんたなんか死ねばいいのに。",
+        "JP",
+    ],
     ["あはははっ！この漫画めっちゃ笑える、見てよこれ、ふふふ、あはは。", "JP"],
-    ["あなたがいなくなって、私は一人になっちゃって、泣いちゃいそうなほど悲しい。", "JP"],
-    ["深層学習の応用により、感情やアクセントを含む声質の微妙な変化も再現されている。", "JP"],
     [
-        "Speech synthesis is the artificial production of human speech.",
-        "EN",
+        "あなたがいなくなって、私は一人になっちゃって、泣いちゃいそうなほど悲しい。",
+        "JP",
+    ],
+    [
+        "深層学習の応用により、感情やアクセントを含む声質の微妙な変化も再現されている。",
+        "JP",
     ],
-    ["语音合成是人工制造人类语音。用于此目的的计算机系统称为语音合成器，可以通过软件或硬件产品实现。", "ZH"],
 ]
-
 initial_md = """
 # Style-Bert-VITS2 JVNVコーパスデモ
-
 怒り・悲しみ・喜び等の感情スタイルを強弱付きで制御できる、[Style-Bert-VITS2](https://github.com/litagin02/Style-Bert-VITS2)のデモです。
-
 入力上限文字数は100文字までにしています。
-
 このデモでは[jvnvのモデル](https://huggingface.co/litagin/style_bert_vits2_jvnv)を使っており、[JVNVコーパス（言語音声と非言語音声を持つ日本語感情音声コーパス）](https://sites.google.com/site/shinnosuketakamichi/research-topics/jvnv_corpus)で学習されたモデルです。
 """
 
-style_md = """
+style_md = f"""
 - プリセットまたは音声ファイルから読み上げの声音・感情・スタイルのようなものを制御できます。
-- デフォルトのNeutralでも、十分に読み上げる文に応じた感情で感情豊かに読み上げられます。このスタイル制御は、それを重み付きで上書きするような感じです。
+- デフォルトの{DEFAULT_STYLE}でも、十分に読み上げる文に応じた感情で感情豊かに読み上げられます。このスタイル制御は、それを重み付きで上書きするような感じです。
 - 強さを大きくしすぎると発音が変になったり声にならなかったりと崩壊することがあります。
 - どのくらいに強さがいい���はモデルやスタイルによって異なるようです。
 - 音声ファイルを入力する場合は、学習データと似た声音の話者（特に同じ性別）でないとよい効果が出ないかもしれません。
@@ -371,7 +198,22 @@ if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("--cpu", action="store_true", help="Use CPU instead of GPU")
     parser.add_argument(
-        "--dir", "-d", type=str, help="Model directory", default=config.out_dir
+        "--dir", "-d", type=str, help="Model directory", default=assets_root
+    )
+    parser.add_argument(
+        "--share", action="store_true", help="Share this app publicly", default=False
+    )
+    parser.add_argument(
+        "--server-name",
+        type=str,
+        default=None,
+        help="Server name for Gradio app",
+    )
+    parser.add_argument(
+        "--no-autolaunch",
+        action="store_true",
+        default=False,
+        help="Do not launch app automatically",
     )
     args = parser.parse_args()
     model_dir = args.dir
@@ -383,12 +225,11 @@ if __name__ == "__main__":
 
     model_holder = ModelHolder(model_dir, device)
 
-    languages = ["JP", "EN", "ZH"]
-    examples = example_hf_spaces if is_hf_spaces else example_local
-
     model_names = model_holder.model_names
     if len(model_names) == 0:
-        logger.error(f"モデルが見つかりませんでした。{model_dir}にモデルを置いてください。")
+        logger.error(
+            f"モデルが見つかりませんでした。{model_dir}にモデルを置いてください。"
+        )
         sys.exit(1)
     initial_id = 0
     initial_pth_files = model_holder.model_files_dict[model_names[initial_id]]
@@ -409,51 +250,87 @@ if __name__ == "__main__":
                             choices=initial_pth_files,
                             value=initial_pth_files[0],
                         )
-                    refresh_button = gr.Button("更新", scale=1, visible=not is_hf_spaces)
+                    refresh_button = gr.Button("更新", scale=1, visible=True)
                     load_button = gr.Button("ロード", scale=1, variant="primary")
                 text_input = gr.TextArea(label="テキスト", value=initial_text)
 
-                line_split = gr.Checkbox(label="改行で分けて生成", value=True)
+                line_split = gr.Checkbox(
+                    label="改行で分けて生成", value=DEFAULT_LINE_SPLIT
+                )
                 split_interval = gr.Slider(
                     minimum=0.0,
                     maximum=2,
-                    value=0.5,
+                    value=DEFAULT_SPLIT_INTERVAL,
                     step=0.1,
-                    label="分けた場合に挟む無音の長さ（秒）",
+                    label="改行ごとに挟む無音の長さ（秒）",
+                )
+                line_split.change(
+                    lambda x: (gr.Slider(visible=x)),
+                    inputs=[line_split],
+                    outputs=[split_interval],
+                )
+                tone = gr.Textbox(
+                    label="アクセント調整（数値は 0=低 か1=高 のみ）",
+                    info="改行で分けない場合のみ使えます。万能ではありません。",
+                )
+                use_tone = gr.Checkbox(label="アクセント調整を使う", value=False)
+                use_tone.change(
+                    lambda x: (gr.Checkbox(value=False) if x else gr.Checkbox()),
+                    inputs=[use_tone],
+                    outputs=[line_split],
                 )
-                language = gr.Dropdown(choices=languages, value="JP", label="Language")
+                language = gr.Dropdown(choices=["JP"], value="JP", label="Language")
+                speaker = gr.Dropdown(label="話者")
                 with gr.Accordion(label="詳細設定", open=False):
                     sdp_ratio = gr.Slider(
-                        minimum=0, maximum=1, value=0.2, step=0.1, label="SDP Ratio"
+                        minimum=0,
+                        maximum=1,
+                        value=DEFAULT_SDP_RATIO,
+                        step=0.1,
+                        label="SDP Ratio",
                     )
                     noise_scale = gr.Slider(
-                        minimum=0.1, maximum=2, value=0.6, step=0.1, label="Noise"
+                        minimum=0.1,
+                        maximum=2,
+                        value=DEFAULT_NOISE,
+                        step=0.1,
+                        label="Noise",
                     )
                     noise_scale_w = gr.Slider(
-                        minimum=0.1, maximum=2, value=0.8, step=0.1, label="Noise_W"
+                        minimum=0.1,
+                        maximum=2,
+                        value=DEFAULT_NOISEW,
+                        step=0.1,
+                        label="Noise_W",
                     )
                     length_scale = gr.Slider(
-                        minimum=0.1, maximum=2, value=1.0, step=0.1, label="Length"
+                        minimum=0.1,
+                        maximum=2,
+                        value=DEFAULT_LENGTH,
+                        step=0.1,
+                        label="Length",
+                    )
+                    use_assist_text = gr.Checkbox(
+                        label="Assist textを使う", value=False
                     )
-                    use_style_text = gr.Checkbox(label="Style textを使う", value=False)
-                    style_text = gr.Textbox(
-                        label="Style text",
+                    assist_text = gr.Textbox(
+                        label="Assist text",
                         placeholder="どうして私の意見を無視するの？許せない、ムカつく！死ねばいいのに。",
                         info="このテキストの読み上げと似た声音・感情になりやすくなります。ただ抑揚やテンポ等が犠牲になる傾向があります。",
                         visible=False,
                     )
-                    style_text_weight = gr.Slider(
+                    assist_text_weight = gr.Slider(
                         minimum=0,
                         maximum=1,
-                        value=0.7,
+                        value=DEFAULT_ASSIST_TEXT_WEIGHT,
                         step=0.1,
-                        label="Style textの強さ",
+                        label="Assist textの強さ",
                         visible=False,
                     )
-                    use_style_text.change(
+                    use_assist_text.change(
                         lambda x: (gr.Textbox(visible=x), gr.Slider(visible=x)),
-                        inputs=[use_style_text],
-                        outputs=[style_text, style_text_weight],
+                        inputs=[use_assist_text],
+                        outputs=[assist_text, assist_text_weight],
                     )
             with gr.Column():
                 with gr.Accordion("スタイルについて詳細", open=False):
@@ -464,25 +341,29 @@ if __name__ == "__main__":
                     value="プリセットから選ぶ",
                 )
                 style = gr.Dropdown(
-                    label="スタイル（Neutralが平均スタイル）",
+                    label=f"スタイル（{DEFAULT_STYLE}が平均スタイル）",
                     choices=["モデルをロードしてください"],
                     value="モデルをロードしてください",
                 )
                 style_weight = gr.Slider(
                     minimum=0,
                     maximum=50,
-                    value=5,
+                    value=DEFAULT_STYLE_WEIGHT,
                     step=0.1,
                     label="スタイルの強さ",
                 )
-                ref_audio_path = gr.Audio(label="参照音声", type="filepath", visible=False)
+                ref_audio_path = gr.Audio(
+                    label="参照音声", type="filepath", visible=False
+                )
                 tts_button = gr.Button(
-                    "音声合成（モデルをロードしてください）", variant="primary", interactive=False
+                    "音声合成（モデルをロードしてください）",
+                    variant="primary",
+                    interactive=False,
                 )
                 text_output = gr.Textbox(label="情報")
                 audio_output = gr.Audio(label="結果")
-                with gr.Accordion("テキスト例", open=False):
-                    gr.Examples(examples, inputs=[text_input, language])
+                with gr.Accordion("テキスト例", open=True):
+                    gr.Examples(example_hf_spaces, inputs=[text_input, language])
 
         tts_button.click(
             tts_fn,
@@ -496,17 +377,20 @@ if __name__ == "__main__":
                 length_scale,
                 line_split,
                 split_interval,
-                style_text,
-                style_text_weight,
-                use_style_text,
+                assist_text,
+                assist_text_weight,
+                use_assist_text,
                 style,
                 style_weight,
+                tone,
+                use_tone,
+                speaker,
             ],
-            outputs=[text_output, audio_output],
+            outputs=[text_output, audio_output, tone],
         )
 
         model_name.change(
-            model_holder.update_model_files_dropdown,
+            model_holder.update_model_files_gr,
             inputs=[model_name],
             outputs=[model_path],
         )
@@ -514,14 +398,14 @@ if __name__ == "__main__":
         model_path.change(make_non_interactive, outputs=[tts_button])
 
         refresh_button.click(
-            model_holder.update_model_names_dropdown,
+            model_holder.update_model_names_gr,
             outputs=[model_name, model_path, tts_button],
         )
 
         load_button.click(
-            model_holder.load_model,
+            model_holder.load_model_gr,
             inputs=[model_name, model_path],
-            outputs=[style, tts_button],
+            outputs=[style, tts_button, speaker],
         )
 
         style_mode.change(
@@ -530,4 +414,6 @@ if __name__ == "__main__":
             outputs=[style, ref_audio_path],
         )
 
-    app.launch(inbrowser=True)
+    app.launch(
+        inbrowser=not args.no_autolaunch, share=args.share, server_name=args.server_name
+    )

attentions.py

@@ -4,7 +4,7 @@ from torch import nn
 from torch.nn import functional as F
 
 import commons
-from tools.log import logger as logging
+from common.log import logger as logging
 
 
 class LayerNorm(nn.Module):

common/constants.py

@@ -0,0 +1,20 @@
+import enum
+
+DEFAULT_STYLE: str = "Neutral"
+DEFAULT_STYLE_WEIGHT: float = 5.0
+
+
+class Languages(str, enum.Enum):
+    JP = "JP"
+    EN = "EN"
+    ZH = "ZH"
+
+
+DEFAULT_SDP_RATIO: float = 0.2
+DEFAULT_NOISE: float = 0.6
+DEFAULT_NOISEW: float = 0.8
+DEFAULT_LENGTH: float = 1.0
+DEFAULT_LINE_SPLIT: bool = True
+DEFAULT_SPLIT_INTERVAL: float = 0.5
+DEFAULT_ASSIST_TEXT_WEIGHT: float = 0.7
+DEFAULT_ASSIST_TEXT_WEIGHT: float = 1.0

common/log.py

@@ -0,0 +1,16 @@
+"""
+logger封装
+"""
+from loguru import logger
+
+from .stdout_wrapper import SAFE_STDOUT
+
+# 移除所有默认的处理器
+logger.remove()
+
+# 自定义格式并添加到标准输出
+log_format = (
+    "<g>{time:MM-DD HH:mm:ss}</g> |<lvl>{level:^8}</lvl>| {file}:{line} | {message}"
+)
+
+logger.add(SAFE_STDOUT, format=log_format, backtrace=True, diagnose=True)

common/stdout_wrapper.py

@@ -0,0 +1,34 @@
+import sys
+import tempfile
+
+
+class StdoutWrapper:
+    def __init__(self):
+        self.temp_file = tempfile.NamedTemporaryFile(mode="w+", delete=False)
+        self.original_stdout = sys.stdout
+
+    def write(self, message: str):
+        self.temp_file.write(message)
+        self.temp_file.flush()
+        print(message, end="", file=self.original_stdout)
+
+    def flush(self):
+        self.temp_file.flush()
+
+    def read(self):
+        self.temp_file.seek(0)
+        return self.temp_file.read()
+
+    def close(self):
+        self.temp_file.close()
+
+    def fileno(self):
+        return self.temp_file.fileno()
+
+
+try:
+    import google.colab
+
+    SAFE_STDOUT = StdoutWrapper()
+except ImportError:
+    SAFE_STDOUT = sys.stdout

common/subprocess_utils.py

@@ -0,0 +1,32 @@
+import subprocess
+import sys
+
+from .log import logger
+from .stdout_wrapper import SAFE_STDOUT
+
+python = sys.executable
+
+
+def run_script_with_log(cmd: list[str], ignore_warning=False) -> tuple[bool, str]:
+    logger.info(f"Running: {' '.join(cmd)}")
+    result = subprocess.run(
+        [python] + cmd,
+        stdout=SAFE_STDOUT,  # type: ignore
+        stderr=subprocess.PIPE,
+        text=True,
+    )
+    if result.returncode != 0:
+        logger.error(f"Error: {' '.join(cmd)}\n{result.stderr}")
+        return False, result.stderr
+    elif result.stderr and not ignore_warning:
+        logger.warning(f"Warning: {' '.join(cmd)}\n{result.stderr}")
+        return True, result.stderr
+    logger.success(f"Success: {' '.join(cmd)}")
+    return True, ""
+
+
+def second_elem_of(original_function):
+    def inner_function(*args, **kwargs):
+        return original_function(*args, **kwargs)[1]
+
+    return inner_function

common/tts_model.py

@@ -0,0 +1,238 @@
+import numpy as np
+import gradio as gr
+import torch
+import os
+import warnings
+from gradio.processing_utils import convert_to_16_bit_wav
+from typing import Dict, List, Optional, Union
+
+import utils
+from infer import get_net_g, infer
+from models import SynthesizerTrn
+from models_jp_extra import SynthesizerTrn as SynthesizerTrnJPExtra
+
+from .log import logger
+from .constants import (
+    DEFAULT_ASSIST_TEXT_WEIGHT,
+    DEFAULT_LENGTH,
+    DEFAULT_LINE_SPLIT,
+    DEFAULT_NOISE,
+    DEFAULT_NOISEW,
+    DEFAULT_SDP_RATIO,
+    DEFAULT_SPLIT_INTERVAL,
+    DEFAULT_STYLE,
+    DEFAULT_STYLE_WEIGHT,
+)
+
+
+class Model:
+    def __init__(
+        self, model_path: str, config_path: str, style_vec_path: str, device: str
+    ):
+        self.model_path: str = model_path
+        self.config_path: str = config_path
+        self.device: str = device
+        self.style_vec_path: str = style_vec_path
+        self.hps: utils.HParams = utils.get_hparams_from_file(self.config_path)
+        self.spk2id: Dict[str, int] = self.hps.data.spk2id
+        self.id2spk: Dict[int, str] = {v: k for k, v in self.spk2id.items()}
+
+        self.num_styles: int = self.hps.data.num_styles
+        if hasattr(self.hps.data, "style2id"):
+            self.style2id: Dict[str, int] = self.hps.data.style2id
+        else:
+            self.style2id: Dict[str, int] = {str(i): i for i in range(self.num_styles)}
+        if len(self.style2id) != self.num_styles:
+            raise ValueError(
+                f"Number of styles ({self.num_styles}) does not match the number of style2id ({len(self.style2id)})"
+            )
+
+        self.style_vectors: np.ndarray = np.load(self.style_vec_path)
+        if self.style_vectors.shape[0] != self.num_styles:
+            raise ValueError(
+                f"The number of styles ({self.num_styles}) does not match the number of style vectors ({self.style_vectors.shape[0]})"
+            )
+
+        self.net_g: Union[SynthesizerTrn, SynthesizerTrnJPExtra, None] = None
+
+    def load_net_g(self):
+        self.net_g = get_net_g(
+            model_path=self.model_path,
+            version=self.hps.version,
+            device=self.device,
+            hps=self.hps,
+        )
+
+    def get_style_vector(self, style_id: int, weight: float = 1.0) -> np.ndarray:
+        mean = self.style_vectors[0]
+        style_vec = self.style_vectors[style_id]
+        style_vec = mean + (style_vec - mean) * weight
+        return style_vec
+
+    def get_style_vector_from_audio(
+        self, audio_path: str, weight: float = 1.0
+    ) -> np.ndarray:
+        from style_gen import get_style_vector
+
+        xvec = get_style_vector(audio_path)
+        mean = self.style_vectors[0]
+        xvec = mean + (xvec - mean) * weight
+        return xvec
+
+    def infer(
+        self,
+        text: str,
+        language: str = "JP",
+        sid: int = 0,
+        reference_audio_path: Optional[str] = None,
+        sdp_ratio: float = DEFAULT_SDP_RATIO,
+        noise: float = DEFAULT_NOISE,
+        noisew: float = DEFAULT_NOISEW,
+        length: float = DEFAULT_LENGTH,
+        line_split: bool = DEFAULT_LINE_SPLIT,
+        split_interval: float = DEFAULT_SPLIT_INTERVAL,
+        assist_text: Optional[str] = None,
+        assist_text_weight: float = DEFAULT_ASSIST_TEXT_WEIGHT,
+        use_assist_text: bool = False,
+        style: str = DEFAULT_STYLE,
+        style_weight: float = DEFAULT_STYLE_WEIGHT,
+        given_tone: Optional[list[int]] = None,
+    ) -> tuple[int, np.ndarray]:
+        logger.info(f"Start generating audio data from text:\n{text}")
+        if language != "JP" and self.hps.version.endswith("JP-Extra"):
+            raise ValueError(
+                "The model is trained with JP-Extra, but the language is not JP"
+            )
+        if reference_audio_path == "":
+            reference_audio_path = None
+        if assist_text == "" or not use_assist_text:
+            assist_text = None
+
+        if self.net_g is None:
+            self.load_net_g()
+        if reference_audio_path is None:
+            style_id = self.style2id[style]
+            style_vector = self.get_style_vector(style_id, style_weight)
+        else:
+            style_vector = self.get_style_vector_from_audio(
+                reference_audio_path, style_weight
+            )
+        if not line_split:
+            with torch.no_grad():
+                audio = infer(
+                    text=text,
+                    sdp_ratio=sdp_ratio,
+                    noise_scale=noise,
+                    noise_scale_w=noisew,
+                    length_scale=length,
+                    sid=sid,
+                    language=language,
+                    hps=self.hps,
+                    net_g=self.net_g,
+                    device=self.device,
+                    assist_text=assist_text,
+                    assist_text_weight=assist_text_weight,
+                    style_vec=style_vector,
+                    given_tone=given_tone,
+                )
+        else:
+            texts = text.split("\n")
+            texts = [t for t in texts if t != ""]
+            audios = []
+            with torch.no_grad():
+                for i, t in enumerate(texts):
+                    audios.append(
+                        infer(
+                            text=t,
+                            sdp_ratio=sdp_ratio,
+                            noise_scale=noise,
+                            noise_scale_w=noisew,
+                            length_scale=length,
+                            sid=sid,
+                            language=language,
+                            hps=self.hps,
+                            net_g=self.net_g,
+                            device=self.device,
+                            assist_text=assist_text,
+                            assist_text_weight=assist_text_weight,
+                            style_vec=style_vector,
+                        )
+                    )
+                    if i != len(texts) - 1:
+                        audios.append(np.zeros(int(44100 * split_interval)))
+                audio = np.concatenate(audios)
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                audio = convert_to_16_bit_wav(audio)
+        logger.info("Audio data generated successfully")
+        return (self.hps.data.sampling_rate, audio)
+
+
+class ModelHolder:
+    def __init__(self, root_dir: str, device: str):
+        self.root_dir: str = root_dir
+        self.device: str = device
+        self.model_files_dict: Dict[str, List[str]] = {}
+        self.current_model: Optional[Model] = None
+        self.model_names: List[str] = []
+        self.models: List[Model] = []
+        self.refresh()
+
+    def refresh(self):
+        self.model_files_dict = {}
+        self.model_names = []
+        self.current_model = None
+        model_dirs = [
+            d
+            for d in os.listdir(self.root_dir)
+            if os.path.isdir(os.path.join(self.root_dir, d))
+        ]
+        for model_name in model_dirs:
+            model_dir = os.path.join(self.root_dir, model_name)
+            model_files = [
+                os.path.join(model_dir, f)
+                for f in os.listdir(model_dir)
+                if f.endswith(".pth") or f.endswith(".pt") or f.endswith(".safetensors")
+            ]
+            if len(model_files) == 0:
+                logger.warning(
+                    f"No model files found in {self.root_dir}/{model_name}, so skip it"
+                )
+                continue
+            self.model_files_dict[model_name] = model_files
+            self.model_names.append(model_name)
+
+    def load_model_gr(
+        self, model_name: str, model_path: str
+    ) -> tuple[gr.Dropdown, gr.Button, gr.Dropdown]:
+        if model_name not in self.model_files_dict:
+            raise ValueError(f"Model `{model_name}` is not found")
+        if model_path not in self.model_files_dict[model_name]:
+            raise ValueError(f"Model file `{model_path}` is not found")
+        self.current_model = Model(
+            model_path=model_path,
+            config_path=os.path.join(self.root_dir, model_name, "config.json"),
+            style_vec_path=os.path.join(self.root_dir, model_name, "style_vectors.npy"),
+            device=self.device,
+        )
+        speakers = list(self.current_model.spk2id.keys())
+        styles = list(self.current_model.style2id.keys())
+        return (
+            gr.Dropdown(choices=styles, value=styles[0]),
+            gr.Button(interactive=True, value="音声合成"),
+            gr.Dropdown(choices=speakers, value=speakers[0]),
+        )
+
+    def update_model_files_gr(self, model_name: str) -> gr.Dropdown:
+        model_files = self.model_files_dict[model_name]
+        return gr.Dropdown(choices=model_files, value=model_files[0])
+
+    def update_model_names_gr(self) -> tuple[gr.Dropdown, gr.Dropdown, gr.Button]:
+        self.refresh()
+        initial_model_name = self.model_names[0]
+        initial_model_files = self.model_files_dict[initial_model_name]
+        return (
+            gr.Dropdown(choices=self.model_names, value=initial_model_name),
+            gr.Dropdown(choices=initial_model_files, value=initial_model_files[0]),
+            gr.Button(interactive=False),  # For tts_button
+        )

config.py

@@ -1,254 +1,269 @@
-"""
-@Desc: 全局配置文件读取
-"""
-import argparse
-import yaml
-from typing import Dict, List
-import os
-import shutil
-import sys
-
-
-class Resample_config:
-    """重采样配置"""
-
-    def __init__(self, in_dir: str, out_dir: str, sampling_rate: int = 44100):
-        self.sampling_rate: int = sampling_rate  # 目标采样率
-        self.in_dir: str = in_dir  # 待处理音频目录路径
-        self.out_dir: str = out_dir  # 重采样输出路径
-
-    @classmethod
-    def from_dict(cls, dataset_path: str, data: Dict[str, any]):
-        """从字典中生成实例"""
-
-        # 不检查路径是否有效，此逻辑在resample.py中处理
-        data["in_dir"] = os.path.join(dataset_path, data["in_dir"])
-        data["out_dir"] = os.path.join(dataset_path, data["out_dir"])
-
-        return cls(**data)
-
-
-class Preprocess_text_config:
-    """数据预处理配置"""
-
-    def __init__(
-        self,
-        transcription_path: str,
-        cleaned_path: str,
-        train_path: str,
-        val_path: str,
-        config_path: str,
-        val_per_lang: int = 5,
-        max_val_total: int = 10000,
-        clean: bool = True,
-    ):
-        self.transcription_path: str = transcription_path  # 原始文本文件路径，文本格式应为{wav_path}|{speaker_name}|{language}|{text}。
-        self.cleaned_path: str = cleaned_path  # 数据清洗后文本路径，可以不填。不填则将在原始文本目录生成
-        self.train_path: str = train_path  # 训练集路径，可以不填。不填则将在原始文本目录生成
-        self.val_path: str = val_path  # 验证集路径，可以不填。不填则将在原始文本目录生成
-        self.config_path: str = config_path  # 配置文件路径
-        self.val_per_lang: int = val_per_lang  # 每个speaker的验证集条数
-        self.max_val_total: int = max_val_total  # 验证集最大条数，多于的会被截断并放到训练集中
-        self.clean: bool = clean  # 是否进行数据清洗
-
-    @classmethod
-    def from_dict(cls, dataset_path: str, data: Dict[str, any]):
-        """从字典中生成实例"""
-
-        data["transcription_path"] = os.path.join(
-            dataset_path, data["transcription_path"]
-        )
-        if data["cleaned_path"] == "" or data["cleaned_path"] is None:
-            data["cleaned_path"] = None
-        else:
-            data["cleaned_path"] = os.path.join(dataset_path, data["cleaned_path"])
-        data["train_path"] = os.path.join(dataset_path, data["train_path"])
-        data["val_path"] = os.path.join(dataset_path, data["val_path"])
-        data["config_path"] = os.path.join(dataset_path, data["config_path"])
-
-        return cls(**data)
-
-
-class Bert_gen_config:
-    """bert_gen 配置"""
-
-    def __init__(
-        self,
-        config_path: str,
-        num_processes: int = 2,
-        device: str = "cuda",
-        use_multi_device: bool = False,
-    ):
-        self.config_path = config_path
-        self.num_processes = num_processes
-        self.device = device
-        self.use_multi_device = use_multi_device
-
-    @classmethod
-    def from_dict(cls, dataset_path: str, data: Dict[str, any]):
-        data["config_path"] = os.path.join(dataset_path, data["config_path"])
-
-        return cls(**data)
-
-
-class Style_gen_config:
-    """style_gen 配置"""
-
-    def __init__(
-        self,
-        config_path: str,
-        num_processes: int = 2,
-        device: str = "cuda",
-    ):
-        self.config_path = config_path
-        self.num_processes = num_processes
-        self.device = device
-
-    @classmethod
-    def from_dict(cls, dataset_path: str, data: Dict[str, any]):
-        data["config_path"] = os.path.join(dataset_path, data["config_path"])
-
-        return cls(**data)
-
-
-class Train_ms_config:
-    """训练配置"""
-
-    def __init__(
-        self,
-        config_path: str,
-        env: Dict[str, any],
-        # base: Dict[str, any],
-        model: str,
-        num_workers: int,
-        spec_cache: bool,
-        keep_ckpts: int,
-    ):
-        self.env = env  # 需要加载的环境变量
-        # self.base = base  # 底模配置
-        self.model = model  # 训练模型存储目录，该路径为相对于dataset_path的路径，而非项目根目录
-        self.config_path = config_path  # 配置文件路径
-        self.num_workers = num_workers  # worker数量
-        self.spec_cache = spec_cache  # 是否启用spec缓存
-        self.keep_ckpts = keep_ckpts  # ckpt数量
-
-    @classmethod
-    def from_dict(cls, dataset_path: str, data: Dict[str, any]):
-        # data["model"] = os.path.join(dataset_path, data["model"])
-        data["config_path"] = os.path.join(dataset_path, data["config_path"])
-
-        return cls(**data)
-
-
-class Webui_config:
-    """webui 配置"""
-
-    def __init__(
-        self,
-        device: str,
-        model: str,
-        config_path: str,
-        language_identification_library: str,
-        port: int = 7860,
-        share: bool = False,
-        debug: bool = False,
-    ):
-        self.device: str = device
-        self.model: str = model  # 端口号
-        self.config_path: str = config_path  # 是否公开部署，对外网开放
-        self.port: int = port  # 是否开启debug模式
-        self.share: bool = share  # 模型路径
-        self.debug: bool = debug  # 配置文件路径
-        self.language_identification_library: str = (
-            language_identification_library  # 语种识别库
-        )
-
-    @classmethod
-    def from_dict(cls, dataset_path: str, data: Dict[str, any]):
-        data["config_path"] = os.path.join(dataset_path, data["config_path"])
-        data["model"] = os.path.join(dataset_path, data["model"])
-        return cls(**data)
-
-
-class Server_config:
-    def __init__(
-        self, models: List[Dict[str, any]], port: int = 5000, device: str = "cuda"
-    ):
-        self.models: List[Dict[str, any]] = models  # 需要加载的所有模型的配置
-        self.port: int = port  # 端口号
-        self.device: str = device  # 模型默认使用设备
-
-    @classmethod
-    def from_dict(cls, data: Dict[str, any]):
-        return cls(**data)
-
-
-class Translate_config:
-    """翻译api配置"""
-
-    def __init__(self, app_key: str, secret_key: str):
-        self.app_key = app_key
-        self.secret_key = secret_key
-
-    @classmethod
-    def from_dict(cls, data: Dict[str, any]):
-        return cls(**data)
-
-
-class Config:
-    def __init__(self, config_path: str):
-        if not os.path.isfile(config_path) and os.path.isfile("default_config.yml"):
-            shutil.copy(src="default_config.yml", dst=config_path)
-            print(
-                f"A configuration file {config_path} has been generated based on the default configuration file default_config.yml."
-            )
-            print(
-                "If you have no special needs, please do not modify default_config.yml."
-            )
-            # sys.exit(0)
-        with open(file=config_path, mode="r", encoding="utf-8") as file:
-            yaml_config: Dict[str, any] = yaml.safe_load(file.read())
-            model_name: str = yaml_config["model_name"]
-            self.model_name: str = model_name
-            if "dataset_path" in yaml_config:
-                dataset_path = yaml_config["dataset_path"]
-            else:
-                dataset_path = f"Data/{model_name}"
-            self.out_dir = yaml_config["out_dir"]
-            # openi_token: str = yaml_config["openi_token"]
-            self.dataset_path: str = dataset_path
-            # self.mirror: str = yaml_config["mirror"]
-            # self.openi_token: str = openi_token
-            self.resample_config: Resample_config = Resample_config.from_dict(
-                dataset_path, yaml_config["resample"]
-            )
-            self.preprocess_text_config: Preprocess_text_config = (
-                Preprocess_text_config.from_dict(
-                    dataset_path, yaml_config["preprocess_text"]
-                )
-            )
-            self.bert_gen_config: Bert_gen_config = Bert_gen_config.from_dict(
-                dataset_path, yaml_config["bert_gen"]
-            )
-            self.style_gen_config: Style_gen_config = Style_gen_config.from_dict(
-                dataset_path, yaml_config["style_gen"]
-            )
-            self.train_ms_config: Train_ms_config = Train_ms_config.from_dict(
-                dataset_path, yaml_config["train_ms"]
-            )
-            self.webui_config: Webui_config = Webui_config.from_dict(
-                dataset_path, yaml_config["webui"]
-            )
-            self.server_config: Server_config = Server_config.from_dict(
-                yaml_config["server"]
-            )
-            # self.translate_config: Translate_config = Translate_config.from_dict(
-            #     yaml_config["translate"]
-            # )
-
-
-parser = argparse.ArgumentParser()
-# 为避免与以前的config.json起冲突，将其更名如下
-parser.add_argument("-y", "--yml_config", type=str, default="config.yml")
-args, _ = parser.parse_known_args()
-config = Config(args.yml_config)
+"""
+@Desc: 全局配置文件读取
+"""
+import argparse
+import os
+import shutil
+from typing import Dict, List
+
+import yaml
+
+from common.log import logger
+
+
+class Resample_config:
+    """重采样配置"""
+
+    def __init__(self, in_dir: str, out_dir: str, sampling_rate: int = 44100):
+        self.sampling_rate: int = sampling_rate  # 目标采样率
+        self.in_dir: str = in_dir  # 待处理音频目录路径
+        self.out_dir: str = out_dir  # 重采样输出路径
+
+    @classmethod
+    def from_dict(cls, dataset_path: str, data: Dict[str, any]):
+        """从字典中生成实例"""
+
+        # 不检查路���是否有效，此逻辑在resample.py中处理
+        data["in_dir"] = os.path.join(dataset_path, data["in_dir"])
+        data["out_dir"] = os.path.join(dataset_path, data["out_dir"])
+
+        return cls(**data)
+
+
+class Preprocess_text_config:
+    """数据预处理配置"""
+
+    def __init__(
+        self,
+        transcription_path: str,
+        cleaned_path: str,
+        train_path: str,
+        val_path: str,
+        config_path: str,
+        val_per_lang: int = 5,
+        max_val_total: int = 10000,
+        clean: bool = True,
+    ):
+        self.transcription_path: str = transcription_path  # 原始文本文件路径，文本格式应为{wav_path}|{speaker_name}|{language}|{text}。
+        self.cleaned_path: str = cleaned_path  # 数据清洗后文本路径，可以不填。不填则将在原始文本目录生成
+        self.train_path: str = train_path  # 训练集路径，可以不填。不填则将在原始文本目录生成
+        self.val_path: str = val_path  # 验证集路径，可以不填。不填则将在原始文本目录生成
+        self.config_path: str = config_path  # 配置文件路径
+        self.val_per_lang: int = val_per_lang  # 每个speaker的验证集条数
+        self.max_val_total: int = max_val_total  # 验证集最大条数，多于的会被截断并放到训练集中
+        self.clean: bool = clean  # 是否进行数据清洗
+
+    @classmethod
+    def from_dict(cls, dataset_path: str, data: Dict[str, any]):
+        """从字典中生成实例"""
+
+        data["transcription_path"] = os.path.join(
+            dataset_path, data["transcription_path"]
+        )
+        if data["cleaned_path"] == "" or data["cleaned_path"] is None:
+            data["cleaned_path"] = None
+        else:
+            data["cleaned_path"] = os.path.join(dataset_path, data["cleaned_path"])
+        data["train_path"] = os.path.join(dataset_path, data["train_path"])
+        data["val_path"] = os.path.join(dataset_path, data["val_path"])
+        data["config_path"] = os.path.join(dataset_path, data["config_path"])
+
+        return cls(**data)
+
+
+class Bert_gen_config:
+    """bert_gen 配置"""
+
+    def __init__(
+        self,
+        config_path: str,
+        num_processes: int = 2,
+        device: str = "cuda",
+        use_multi_device: bool = False,
+    ):
+        self.config_path = config_path
+        self.num_processes = num_processes
+        self.device = device
+        self.use_multi_device = use_multi_device
+
+    @classmethod
+    def from_dict(cls, dataset_path: str, data: Dict[str, any]):
+        data["config_path"] = os.path.join(dataset_path, data["config_path"])
+
+        return cls(**data)
+
+
+class Style_gen_config:
+    """style_gen 配置"""
+
+    def __init__(
+        self,
+        config_path: str,
+        num_processes: int = 4,
+        device: str = "cuda",
+    ):
+        self.config_path = config_path
+        self.num_processes = num_processes
+        self.device = device
+
+    @classmethod
+    def from_dict(cls, dataset_path: str, data: Dict[str, any]):
+        data["config_path"] = os.path.join(dataset_path, data["config_path"])
+
+        return cls(**data)
+
+
+class Train_ms_config:
+    """训练配置"""
+
+    def __init__(
+        self,
+        config_path: str,
+        env: Dict[str, any],
+        # base: Dict[str, any],
+        model_dir: str,
+        num_workers: int,
+        spec_cache: bool,
+        keep_ckpts: int,
+    ):
+        self.env = env  # 需要加载的环境变量
+        # self.base = base  # 底模配置
+        self.model_dir = model_dir  # 训练模型存储目录，该路径为相对于dataset_path的路径，而非项目根目录
+        self.config_path = config_path  # 配置文件路径
+        self.num_workers = num_workers  # worker数量
+        self.spec_cache = spec_cache  # 是否启用spec缓存
+        self.keep_ckpts = keep_ckpts  # ckpt数量
+
+    @classmethod
+    def from_dict(cls, dataset_path: str, data: Dict[str, any]):
+        # data["model"] = os.path.join(dataset_path, data["model"])
+        data["config_path"] = os.path.join(dataset_path, data["config_path"])
+
+        return cls(**data)
+
+
+class Webui_config:
+    """webui 配置"""
+
+    def __init__(
+        self,
+        device: str,
+        model: str,
+        config_path: str,
+        language_identification_library: str,
+        port: int = 7860,
+        share: bool = False,
+        debug: bool = False,
+    ):
+        self.device: str = device
+        self.model: str = model  # 端口号
+        self.config_path: str = config_path  # 是否公开部署，对外网开放
+        self.port: int = port  # 是否开启debug模式
+        self.share: bool = share  # 模型路径
+        self.debug: bool = debug  # 配置文件路径
+        self.language_identification_library: str = (
+            language_identification_library  # 语种识别库
+        )
+
+    @classmethod
+    def from_dict(cls, dataset_path: str, data: Dict[str, any]):
+        data["config_path"] = os.path.join(dataset_path, data["config_path"])
+        data["model"] = os.path.join(dataset_path, data["model"])
+        return cls(**data)
+
+
+class Server_config:
+    def __init__(
+        self,
+        port: int = 5000,
+        device: str = "cuda",
+        limit: int = 100,
+        language: str = "JP",
+        origins: List[str] = None,
+    ):
+        self.port: int = port
+        self.device: str = device
+        self.language: str = language
+        self.limit: int = limit
+        self.origins: List[str] = origins
+
+    @classmethod
+    def from_dict(cls, data: Dict[str, any]):
+        return cls(**data)
+
+
+class Translate_config:
+    """翻译api配置"""
+
+    def __init__(self, app_key: str, secret_key: str):
+        self.app_key = app_key
+        self.secret_key = secret_key
+
+    @classmethod
+    def from_dict(cls, data: Dict[str, any]):
+        return cls(**data)
+
+
+class Config:
+    def __init__(self, config_path: str, path_config: dict[str, str]):
+        if not os.path.isfile(config_path) and os.path.isfile("default_config.yml"):
+            shutil.copy(src="default_config.yml", dst=config_path)
+            logger.info(
+                f"A configuration file {config_path} has been generated based on the default configuration file default_config.yml."
+            )
+            logger.info(
+                "If you have no special needs, please do not modify default_config.yml."
+            )
+            # sys.exit(0)
+        with open(file=config_path, mode="r", encoding="utf-8") as file:
+            yaml_config: Dict[str, any] = yaml.safe_load(file.read())
+            model_name: str = yaml_config["model_name"]
+            self.model_name: str = model_name
+            if "dataset_path" in yaml_config:
+                dataset_path = yaml_config["dataset_path"]
+            else:
+                dataset_path = os.path.join(path_config["dataset_root"], model_name)
+            self.dataset_path: str = dataset_path
+            self.assets_root: str = path_config["assets_root"]
+            self.out_dir = os.path.join(self.assets_root, model_name)
+            self.resample_config: Resample_config = Resample_config.from_dict(
+                dataset_path, yaml_config["resample"]
+            )
+            self.preprocess_text_config: Preprocess_text_config = (
+                Preprocess_text_config.from_dict(
+                    dataset_path, yaml_config["preprocess_text"]
+                )
+            )
+            self.bert_gen_config: Bert_gen_config = Bert_gen_config.from_dict(
+                dataset_path, yaml_config["bert_gen"]
+            )
+            self.style_gen_config: Style_gen_config = Style_gen_config.from_dict(
+                dataset_path, yaml_config["style_gen"]
+            )
+            self.train_ms_config: Train_ms_config = Train_ms_config.from_dict(
+                dataset_path, yaml_config["train_ms"]
+            )
+            self.webui_config: Webui_config = Webui_config.from_dict(
+                dataset_path, yaml_config["webui"]
+            )
+            self.server_config: Server_config = Server_config.from_dict(
+                yaml_config["server"]
+            )
+            # self.translate_config: Translate_config = Translate_config.from_dict(
+            #     yaml_config["translate"]
+            # )
+
+
+with open(os.path.join("configs", "paths.yml"), "r", encoding="utf-8") as f:
+    path_config: dict[str, str] = yaml.safe_load(f.read())
+    # Should contain the following keys:
+    # - dataset_root: the root directory of the dataset, default to "Data"
+    # - assets_root: the root directory of the assets, default to "model_assets"
+
+
+try:
+    config = Config("config.yml", path_config)
+except (TypeError, KeyError):
+    logger.warning("Old config.yml found. Replace it with default_config.yml.")
+    shutil.copy(src="default_config.yml", dst="config.yml")
+    config = Config("config.yml", path_config)

config.yml

@@ -1,58 +1,51 @@
-bert_gen:
-  config_path: config.json
-  device: cuda
-  num_processes: 4
-  use_multi_device: false
-dataset_path: Data\jvnv-M2
-model_name: jvnv-M2
-out_dir: model_assets
-preprocess_text:
-  clean: true
-  cleaned_path: ''
-  config_path: config.json
-  max_val_total: 12
-  train_path: filelists/train.list
-  transcription_path: filelists/text.list
-  val_path: filelists/val.list
-  val_per_lang: 4
-resample:
-  in_dir: audios/raw
-  out_dir: audios/wavs
-  sampling_rate: 44100
-server:
-  device: cuda
-  models:
-  - config: ''
-    device: cuda
-    language: ZH
-    model: ''
-  - config: ''
-    device: cpu
-    language: JP
-    model: ''
-    speakers: []
-  port: 5000
-style_gen:
-  config_path: config.json
-  device: cuda
-  num_processes: 4
-train_ms:
-  config_path: config.json
-  env:
-    LOCAL_RANK: 0
-    MASTER_ADDR: localhost
-    MASTER_PORT: 10086
-    RANK: 0
-    WORLD_SIZE: 1
-  keep_ckpts: 1
-  model: models
-  num_workers: 16
-  spec_cache: true
-webui:
-  config_path: config.json
-  debug: false
-  device: cuda
-  language_identification_library: langid
-  model: models/G_8000.pth
-  port: 7860
-  share: false
+bert_gen:
+  config_path: config.json
+  device: cuda
+  num_processes: 2
+  use_multi_device: false
+dataset_path: Data\model_name
+model_name: model_name
+preprocess_text:
+  clean: true
+  cleaned_path: ''
+  config_path: config.json
+  max_val_total: 12
+  train_path: train.list
+  transcription_path: esd.list
+  val_path: val.list
+  val_per_lang: 4
+resample:
+  in_dir: raw
+  out_dir: wavs
+  sampling_rate: 44100
+server:
+  device: cuda
+  language: JP
+  limit: 100
+  origins:
+  - '*'
+  port: 5000
+style_gen:
+  config_path: config.json
+  device: cuda
+  num_processes: 4
+train_ms:
+  config_path: config.json
+  env:
+    LOCAL_RANK: 0
+    MASTER_ADDR: localhost
+    MASTER_PORT: 10086
+    RANK: 0
+    WORLD_SIZE: 1
+  keep_ckpts: 1
+  model_dir: models
+  num_workers: 16
+  spec_cache: true
+webui:
+  config_path: config.json
+  debug: false
+  device: cuda
+  language_identification_library: langid
+  model: models/G_8000.pth
+  port: 7860
+  share: false

configs/config.json

@@ -0,0 +1,71 @@
+{
+  "model_name": "your_model_name",
+  "train": {
+    "log_interval": 200,
+    "eval_interval": 1000,
+    "seed": 42,
+    "epochs": 1000,
+    "learning_rate": 0.0002,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 4,
+    "bf16_run": true,
+    "lr_decay": 0.99995,
+    "segment_size": 16384,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0,
+    "skip_optimizer": false,
+    "freeze_ZH_bert": false,
+    "freeze_JP_bert": false,
+    "freeze_EN_bert": false,
+    "freeze_style": false
+  },
+  "data": {
+    "training_files": "Data/your_model_name/filelists/train.list",
+    "validation_files": "Data/your_model_name/filelists/val.list",
+    "max_wav_value": 32768.0,
+    "sampling_rate": 44100,
+    "filter_length": 2048,
+    "hop_length": 512,
+    "win_length": 2048,
+    "n_mel_channels": 128,
+    "mel_fmin": 0.0,
+    "mel_fmax": null,
+    "add_blank": true,
+    "n_speakers": 1,
+    "cleaned_text": true,
+    "num_styles": 1,
+    "style2id": {
+      "Neutral": 0
+    }
+  },
+  "model": {
+    "use_spk_conditioned_encoder": true,
+    "use_noise_scaled_mas": true,
+    "use_mel_posterior_encoder": false,
+    "use_duration_discriminator": true,
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3, 7, 11],
+    "resblock_dilation_sizes": [
+      [1, 3, 5],
+      [1, 3, 5],
+      [1, 3, 5]
+    ],
+    "upsample_rates": [8, 8, 2, 2, 2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [16, 16, 8, 2, 2],
+    "n_layers_q": 3,
+    "use_spectral_norm": false,
+    "gin_channels": 256
+  },
+  "version": "2.0.1"
+}

configs/configs_jp_extra.json

@@ -0,0 +1,78 @@
+{
+  "train": {
+    "log_interval": 200,
+    "eval_interval": 1000,
+    "seed": 42,
+    "epochs": 1000,
+    "learning_rate": 0.0001,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 24,
+    "bf16_run": false,
+    "fp16_run": false,
+    "lr_decay": 0.99996,
+    "segment_size": 16384,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0,
+    "c_commit": 100,
+    "skip_optimizer": true,
+    "freeze_ZH_bert": false,
+    "freeze_JP_bert": false,
+    "freeze_EN_bert": false,
+    "freeze_emo": false,
+    "freeze_style": false
+  },
+  "data": {
+    "use_jp_extra": true,
+    "training_files": "filelists/train.list",
+    "validation_files": "filelists/val.list",
+    "max_wav_value": 32768.0,
+    "sampling_rate": 44100,
+    "filter_length": 2048,
+    "hop_length": 512,
+    "win_length": 2048,
+    "n_mel_channels": 128,
+    "mel_fmin": 0.0,
+    "mel_fmax": null,
+    "add_blank": true,
+    "n_speakers": 512,
+    "cleaned_text": true
+  },
+  "model": {
+    "use_spk_conditioned_encoder": true,
+    "use_noise_scaled_mas": true,
+    "use_mel_posterior_encoder": false,
+    "use_duration_discriminator": false,
+    "use_wavlm_discriminator": true,
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3, 7, 11],
+    "resblock_dilation_sizes": [
+      [1, 3, 5],
+      [1, 3, 5],
+      [1, 3, 5]
+    ],
+    "upsample_rates": [8, 8, 2, 2, 2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [16, 16, 8, 2, 2],
+    "n_layers_q": 3,
+    "use_spectral_norm": false,
+    "gin_channels": 512,
+    "slm": {
+      "model": "./slm/wavlm-base-plus",
+      "sr": 16000,
+      "hidden": 768,
+      "nlayers": 13,
+      "initial_channel": 64
+    }
+  },
+  "version": "2.0.1-JP-Extra"
+}

configs/paths.yml

@@ -0,0 +1,8 @@
+# Root directory of the training dataset.
+# The training dataset of {model_name} should be placed in {dataset_root}/{model_name}.
+dataset_root: Data
+
+# Root directory of the model assets (for inference).
+# In training, the model assets will be saved to {assets_root}/{model_name},
+# and in inference, we load all the models from {assets_root}.
+assets_root: model_assets

infer.py

@@ -1,263 +1,306 @@
-import torch
-
-import commons
-import utils
-from models import SynthesizerTrn
-from text import cleaned_text_to_sequence, get_bert
-from text.cleaner import clean_text
-from text.symbols import symbols
-
-# latest_version = "1.0"
-
-
-def get_net_g(model_path: str, version: str, device: str, hps):
-    net_g = SynthesizerTrn(
-        len(symbols),
-        hps.data.filter_length // 2 + 1,
-        hps.train.segment_size // hps.data.hop_length,
-        n_speakers=hps.data.n_speakers,
-        **hps.model,
-    ).to(device)
-    net_g.state_dict()
-    _ = net_g.eval()
-    if model_path.endswith(".pth") or model_path.endswith(".pt"):
-        _ = utils.load_checkpoint(model_path, net_g, None, skip_optimizer=True)
-    elif model_path.endswith(".safetensors"):
-        _ = utils.load_safetensors(model_path, net_g, device)
-    else:
-        raise ValueError(f"Unknown model format: {model_path}")
-    return net_g
-
-
-def get_text(text, language_str, hps, device, style_text=None, style_weight=0.7):
-    # 在此处实现当前版本的get_text
-    norm_text, phone, tone, word2ph = clean_text(text, language_str)
-    phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str)
-
-    if hps.data.add_blank:
-        phone = commons.intersperse(phone, 0)
-        tone = commons.intersperse(tone, 0)
-        language = commons.intersperse(language, 0)
-        for i in range(len(word2ph)):
-            word2ph[i] = word2ph[i] * 2
-        word2ph[0] += 1
-    bert_ori = get_bert(
-        norm_text, word2ph, language_str, device, style_text, style_weight
-    )
-    del word2ph
-    assert bert_ori.shape[-1] == len(phone), phone
-
-    if language_str == "ZH":
-        bert = bert_ori
-        ja_bert = torch.zeros(1024, len(phone))
-        en_bert = torch.zeros(1024, len(phone))
-    elif language_str == "JP":
-        bert = torch.zeros(1024, len(phone))
-        ja_bert = bert_ori
-        en_bert = torch.zeros(1024, len(phone))
-    elif language_str == "EN":
-        bert = torch.zeros(1024, len(phone))
-        ja_bert = torch.zeros(1024, len(phone))
-        en_bert = bert_ori
-    else:
-        raise ValueError("language_str should be ZH, JP or EN")
-
-    assert bert.shape[-1] == len(
-        phone
-    ), f"Bert seq len {bert.shape[-1]} != {len(phone)}"
-
-    phone = torch.LongTensor(phone)
-    tone = torch.LongTensor(tone)
-    language = torch.LongTensor(language)
-    return bert, ja_bert, en_bert, phone, tone, language
-
-
-def infer(
-    text,
-    style_vec,
-    sdp_ratio,
-    noise_scale,
-    noise_scale_w,
-    length_scale,
-    sid: int,  # In the original Bert-VITS2, its speaker_name: str, but here it's id
-    language,
-    hps,
-    net_g,
-    device,
-    skip_start=False,
-    skip_end=False,
-    style_text=None,
-    style_weight=0.7,
-):
-    bert, ja_bert, en_bert, phones, tones, lang_ids = get_text(
-        text,
-        language,
-        hps,
-        device,
-        style_text=style_text,
-        style_weight=style_weight,
-    )
-    if skip_start:
-        phones = phones[3:]
-        tones = tones[3:]
-        lang_ids = lang_ids[3:]
-        bert = bert[:, 3:]
-        ja_bert = ja_bert[:, 3:]
-        en_bert = en_bert[:, 3:]
-    if skip_end:
-        phones = phones[:-2]
-        tones = tones[:-2]
-        lang_ids = lang_ids[:-2]
-        bert = bert[:, :-2]
-        ja_bert = ja_bert[:, :-2]
-        en_bert = en_bert[:, :-2]
-    with torch.no_grad():
-        x_tst = phones.to(device).unsqueeze(0)
-        tones = tones.to(device).unsqueeze(0)
-        lang_ids = lang_ids.to(device).unsqueeze(0)
-        bert = bert.to(device).unsqueeze(0)
-        ja_bert = ja_bert.to(device).unsqueeze(0)
-        en_bert = en_bert.to(device).unsqueeze(0)
-        x_tst_lengths = torch.LongTensor([phones.size(0)]).to(device)
-        style_vec = torch.from_numpy(style_vec).to(device).unsqueeze(0)
-        del phones
-        sid_tensor = torch.LongTensor([sid]).to(device)
-        audio = (
-            net_g.infer(
-                x_tst,
-                x_tst_lengths,
-                sid_tensor,
-                tones,
-                lang_ids,
-                bert,
-                ja_bert,
-                en_bert,
-                style_vec=style_vec,
-                sdp_ratio=sdp_ratio,
-                noise_scale=noise_scale,
-                noise_scale_w=noise_scale_w,
-                length_scale=length_scale,
-            )[0][0, 0]
-            .data.cpu()
-            .float()
-            .numpy()
-        )
-        del (
-            x_tst,
-            tones,
-            lang_ids,
-            bert,
-            x_tst_lengths,
-            sid_tensor,
-            ja_bert,
-            en_bert,
-            style_vec,
-        )  # , emo
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-        return audio
-
-
-def infer_multilang(
-    text,
-    style_vec,
-    sdp_ratio,
-    noise_scale,
-    noise_scale_w,
-    length_scale,
-    sid,
-    language,
-    hps,
-    net_g,
-    device,
-    skip_start=False,
-    skip_end=False,
-):
-    bert, ja_bert, en_bert, phones, tones, lang_ids = [], [], [], [], [], []
-    # emo = get_emo_(reference_audio, emotion, sid)
-    # if isinstance(reference_audio, np.ndarray):
-    #     emo = get_clap_audio_feature(reference_audio, device)
-    # else:
-    #     emo = get_clap_text_feature(emotion, device)
-    # emo = torch.squeeze(emo, dim=1)
-    for idx, (txt, lang) in enumerate(zip(text, language)):
-        _skip_start = (idx != 0) or (skip_start and idx == 0)
-        _skip_end = (idx != len(language) - 1) or skip_end
-        (
-            temp_bert,
-            temp_ja_bert,
-            temp_en_bert,
-            temp_phones,
-            temp_tones,
-            temp_lang_ids,
-        ) = get_text(txt, lang, hps, device)
-        if _skip_start:
-            temp_bert = temp_bert[:, 3:]
-            temp_ja_bert = temp_ja_bert[:, 3:]
-            temp_en_bert = temp_en_bert[:, 3:]
-            temp_phones = temp_phones[3:]
-            temp_tones = temp_tones[3:]
-            temp_lang_ids = temp_lang_ids[3:]
-        if _skip_end:
-            temp_bert = temp_bert[:, :-2]
-            temp_ja_bert = temp_ja_bert[:, :-2]
-            temp_en_bert = temp_en_bert[:, :-2]
-            temp_phones = temp_phones[:-2]
-            temp_tones = temp_tones[:-2]
-            temp_lang_ids = temp_lang_ids[:-2]
-        bert.append(temp_bert)
-        ja_bert.append(temp_ja_bert)
-        en_bert.append(temp_en_bert)
-        phones.append(temp_phones)
-        tones.append(temp_tones)
-        lang_ids.append(temp_lang_ids)
-    bert = torch.concatenate(bert, dim=1)
-    ja_bert = torch.concatenate(ja_bert, dim=1)
-    en_bert = torch.concatenate(en_bert, dim=1)
-    phones = torch.concatenate(phones, dim=0)
-    tones = torch.concatenate(tones, dim=0)
-    lang_ids = torch.concatenate(lang_ids, dim=0)
-    with torch.no_grad():
-        x_tst = phones.to(device).unsqueeze(0)
-        tones = tones.to(device).unsqueeze(0)
-        lang_ids = lang_ids.to(device).unsqueeze(0)
-        bert = bert.to(device).unsqueeze(0)
-        ja_bert = ja_bert.to(device).unsqueeze(0)
-        en_bert = en_bert.to(device).unsqueeze(0)
-        # emo = emo.to(device).unsqueeze(0)
-        x_tst_lengths = torch.LongTensor([phones.size(0)]).to(device)
-        del phones
-        speakers = torch.LongTensor([hps.data.spk2id[sid]]).to(device)
-        audio = (
-            net_g.infer(
-                x_tst,
-                x_tst_lengths,
-                speakers,
-                tones,
-                lang_ids,
-                bert,
-                ja_bert,
-                en_bert,
-                style_vec=style_vec,
-                sdp_ratio=sdp_ratio,
-                noise_scale=noise_scale,
-                noise_scale_w=noise_scale_w,
-                length_scale=length_scale,
-            )[0][0, 0]
-            .data.cpu()
-            .float()
-            .numpy()
-        )
-        del (
-            x_tst,
-            tones,
-            lang_ids,
-            bert,
-            x_tst_lengths,
-            speakers,
-            ja_bert,
-            en_bert,
-        )  # , emo
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-        return audio
+import torch
+
+import commons
+import utils
+from models import SynthesizerTrn
+from models_jp_extra import SynthesizerTrn as SynthesizerTrnJPExtra
+from text import cleaned_text_to_sequence, get_bert
+from text.cleaner import clean_text
+from text.symbols import symbols
+from common.log import logger
+
+
+class InvalidToneError(ValueError):
+    pass
+
+
+def get_net_g(model_path: str, version: str, device: str, hps):
+    if version.endswith("JP-Extra"):
+        logger.info("Using JP-Extra model")
+        net_g = SynthesizerTrnJPExtra(
+            len(symbols),
+            hps.data.filter_length // 2 + 1,
+            hps.train.segment_size // hps.data.hop_length,
+            n_speakers=hps.data.n_speakers,
+            **hps.model,
+        ).to(device)
+    else:
+        logger.info("Using normal model")
+        net_g = SynthesizerTrn(
+            len(symbols),
+            hps.data.filter_length // 2 + 1,
+            hps.train.segment_size // hps.data.hop_length,
+            n_speakers=hps.data.n_speakers,
+            **hps.model,
+        ).to(device)
+    net_g.state_dict()
+    _ = net_g.eval()
+    if model_path.endswith(".pth") or model_path.endswith(".pt"):
+        _ = utils.load_checkpoint(model_path, net_g, None, skip_optimizer=True)
+    elif model_path.endswith(".safetensors"):
+        _ = utils.load_safetensors(model_path, net_g, True)
+    else:
+        raise ValueError(f"Unknown model format: {model_path}")
+    return net_g
+
+
+def get_text(
+    text,
+    language_str,
+    hps,
+    device,
+    assist_text=None,
+    assist_text_weight=0.7,
+    given_tone=None,
+):
+    use_jp_extra = hps.version.endswith("JP-Extra")
+    norm_text, phone, tone, word2ph = clean_text(text, language_str, use_jp_extra)
+    if given_tone is not None:
+        if len(given_tone) != len(phone):
+            raise InvalidToneError(
+                f"Length of given_tone ({len(given_tone)}) != length of phone ({len(phone)})"
+            )
+        tone = given_tone
+    phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str)
+
+    if hps.data.add_blank:
+        phone = commons.intersperse(phone, 0)
+        tone = commons.intersperse(tone, 0)
+        language = commons.intersperse(language, 0)
+        for i in range(len(word2ph)):
+            word2ph[i] = word2ph[i] * 2
+        word2ph[0] += 1
+    bert_ori = get_bert(
+        norm_text, word2ph, language_str, device, assist_text, assist_text_weight
+    )
+    del word2ph
+    assert bert_ori.shape[-1] == len(phone), phone
+
+    if language_str == "ZH":
+        bert = bert_ori
+        ja_bert = torch.zeros(1024, len(phone))
+        en_bert = torch.zeros(1024, len(phone))
+    elif language_str == "JP":
+        bert = torch.zeros(1024, len(phone))
+        ja_bert = bert_ori
+        en_bert = torch.zeros(1024, len(phone))
+    elif language_str == "EN":
+        bert = torch.zeros(1024, len(phone))
+        ja_bert = torch.zeros(1024, len(phone))
+        en_bert = bert_ori
+    else:
+        raise ValueError("language_str should be ZH, JP or EN")
+
+    assert bert.shape[-1] == len(
+        phone
+    ), f"Bert seq len {bert.shape[-1]} != {len(phone)}"
+
+    phone = torch.LongTensor(phone)
+    tone = torch.LongTensor(tone)
+    language = torch.LongTensor(language)
+    return bert, ja_bert, en_bert, phone, tone, language
+
+
+def infer(
+    text,
+    style_vec,
+    sdp_ratio,
+    noise_scale,
+    noise_scale_w,
+    length_scale,
+    sid: int,  # In the original Bert-VITS2, its speaker_name: str, but here it's id
+    language,
+    hps,
+    net_g,
+    device,
+    skip_start=False,
+    skip_end=False,
+    assist_text=None,
+    assist_text_weight=0.7,
+    given_tone=None,
+):
+    is_jp_extra = hps.version.endswith("JP-Extra")
+    bert, ja_bert, en_bert, phones, tones, lang_ids = get_text(
+        text,
+        language,
+        hps,
+        device,
+        assist_text=assist_text,
+        assist_text_weight=assist_text_weight,
+        given_tone=given_tone,
+    )
+    if skip_start:
+        phones = phones[3:]
+        tones = tones[3:]
+        lang_ids = lang_ids[3:]
+        bert = bert[:, 3:]
+        ja_bert = ja_bert[:, 3:]
+        en_bert = en_bert[:, 3:]
+    if skip_end:
+        phones = phones[:-2]
+        tones = tones[:-2]
+        lang_ids = lang_ids[:-2]
+        bert = bert[:, :-2]
+        ja_bert = ja_bert[:, :-2]
+        en_bert = en_bert[:, :-2]
+    with torch.no_grad():
+        x_tst = phones.to(device).unsqueeze(0)
+        tones = tones.to(device).unsqueeze(0)
+        lang_ids = lang_ids.to(device).unsqueeze(0)
+        bert = bert.to(device).unsqueeze(0)
+        ja_bert = ja_bert.to(device).unsqueeze(0)
+        en_bert = en_bert.to(device).unsqueeze(0)
+        x_tst_lengths = torch.LongTensor([phones.size(0)]).to(device)
+        style_vec = torch.from_numpy(style_vec).to(device).unsqueeze(0)
+        del phones
+        sid_tensor = torch.LongTensor([sid]).to(device)
+        if is_jp_extra:
+            output = net_g.infer(
+                x_tst,
+                x_tst_lengths,
+                sid_tensor,
+                tones,
+                lang_ids,
+                ja_bert,
+                style_vec=style_vec,
+                sdp_ratio=sdp_ratio,
+                noise_scale=noise_scale,
+                noise_scale_w=noise_scale_w,
+                length_scale=length_scale,
+            )
+        else:
+            output = net_g.infer(
+                x_tst,
+                x_tst_lengths,
+                sid_tensor,
+                tones,
+                lang_ids,
+                bert,
+                ja_bert,
+                en_bert,
+                style_vec=style_vec,
+                sdp_ratio=sdp_ratio,
+                noise_scale=noise_scale,
+                noise_scale_w=noise_scale_w,
+                length_scale=length_scale,
+            )
+        audio = output[0][0, 0].data.cpu().float().numpy()
+        del (
+            x_tst,
+            tones,
+            lang_ids,
+            bert,
+            x_tst_lengths,
+            sid_tensor,
+            ja_bert,
+            en_bert,
+            style_vec,
+        )  # , emo
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        return audio
+
+
+def infer_multilang(
+    text,
+    style_vec,
+    sdp_ratio,
+    noise_scale,
+    noise_scale_w,
+    length_scale,
+    sid,
+    language,
+    hps,
+    net_g,
+    device,
+    skip_start=False,
+    skip_end=False,
+):
+    bert, ja_bert, en_bert, phones, tones, lang_ids = [], [], [], [], [], []
+    # emo = get_emo_(reference_audio, emotion, sid)
+    # if isinstance(reference_audio, np.ndarray):
+    #     emo = get_clap_audio_feature(reference_audio, device)
+    # else:
+    #     emo = get_clap_text_feature(emotion, device)
+    # emo = torch.squeeze(emo, dim=1)
+    for idx, (txt, lang) in enumerate(zip(text, language)):
+        _skip_start = (idx != 0) or (skip_start and idx == 0)
+        _skip_end = (idx != len(language) - 1) or skip_end
+        (
+            temp_bert,
+            temp_ja_bert,
+            temp_en_bert,
+            temp_phones,
+            temp_tones,
+            temp_lang_ids,
+        ) = get_text(txt, lang, hps, device)
+        if _skip_start:
+            temp_bert = temp_bert[:, 3:]
+            temp_ja_bert = temp_ja_bert[:, 3:]
+            temp_en_bert = temp_en_bert[:, 3:]
+            temp_phones = temp_phones[3:]
+            temp_tones = temp_tones[3:]
+            temp_lang_ids = temp_lang_ids[3:]
+        if _skip_end:
+            temp_bert = temp_bert[:, :-2]
+            temp_ja_bert = temp_ja_bert[:, :-2]
+            temp_en_bert = temp_en_bert[:, :-2]
+            temp_phones = temp_phones[:-2]
+            temp_tones = temp_tones[:-2]
+            temp_lang_ids = temp_lang_ids[:-2]
+        bert.append(temp_bert)
+        ja_bert.append(temp_ja_bert)
+        en_bert.append(temp_en_bert)
+        phones.append(temp_phones)
+        tones.append(temp_tones)
+        lang_ids.append(temp_lang_ids)
+    bert = torch.concatenate(bert, dim=1)
+    ja_bert = torch.concatenate(ja_bert, dim=1)
+    en_bert = torch.concatenate(en_bert, dim=1)
+    phones = torch.concatenate(phones, dim=0)
+    tones = torch.concatenate(tones, dim=0)
+    lang_ids = torch.concatenate(lang_ids, dim=0)
+    with torch.no_grad():
+        x_tst = phones.to(device).unsqueeze(0)
+        tones = tones.to(device).unsqueeze(0)
+        lang_ids = lang_ids.to(device).unsqueeze(0)
+        bert = bert.to(device).unsqueeze(0)
+        ja_bert = ja_bert.to(device).unsqueeze(0)
+        en_bert = en_bert.to(device).unsqueeze(0)
+        # emo = emo.to(device).unsqueeze(0)
+        x_tst_lengths = torch.LongTensor([phones.size(0)]).to(device)
+        del phones
+        speakers = torch.LongTensor([hps.data.spk2id[sid]]).to(device)
+        audio = (
+            net_g.infer(
+                x_tst,
+                x_tst_lengths,
+                speakers,
+                tones,
+                lang_ids,
+                bert,
+                ja_bert,
+                en_bert,
+                style_vec=style_vec,
+                sdp_ratio=sdp_ratio,
+                noise_scale=noise_scale,
+                noise_scale_w=noise_scale_w,
+                length_scale=length_scale,
+            )[0][0, 0]
+            .data.cpu()
+            .float()
+            .numpy()
+        )
+        del (
+            x_tst,
+            tones,
+            lang_ids,
+            bert,
+            x_tst_lengths,
+            speakers,
+            ja_bert,
+            en_bert,
+        )  # , emo
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        return audio

model_assets/jvnv-F1-jp/config.json

@@ -0,0 +1,92 @@
+{
+  "train": {
+    "log_interval": 200,
+    "eval_interval": 1000,
+    "seed": 42,
+    "epochs": 300,
+    "learning_rate": 0.0001,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 4,
+    "bf16_run": true,
+    "fp16_run": false,
+    "lr_decay": 0.99996,
+    "segment_size": 16384,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0,
+    "c_commit": 100,
+    "skip_optimizer": true,
+    "freeze_ZH_bert": false,
+    "freeze_JP_bert": false,
+    "freeze_EN_bert": false,
+    "freeze_emo": false,
+    "freeze_style": false
+  },
+  "data": {
+    "use_jp_extra": true,
+    "training_files": "Data/jvnv-F1-jp/train.list",
+    "validation_files": "Data/jvnv-F1-jp/val.list",
+    "max_wav_value": 32768.0,
+    "sampling_rate": 44100,
+    "filter_length": 2048,
+    "hop_length": 512,
+    "win_length": 2048,
+    "n_mel_channels": 128,
+    "mel_fmin": 0.0,
+    "mel_fmax": null,
+    "add_blank": true,
+    "n_speakers": 1,
+    "cleaned_text": true,
+    "spk2id": {
+      "jvnv-F1-jp": 0
+    },
+    "num_styles": 7,
+    "style2id": {
+      "Neutral": 0,
+      "Angry": 1,
+      "Disgust": 2,
+      "Fear": 3,
+      "Happy": 4,
+      "Sad": 5,
+      "Surprise": 6
+    }
+  },
+  "model": {
+    "use_spk_conditioned_encoder": true,
+    "use_noise_scaled_mas": true,
+    "use_mel_posterior_encoder": false,
+    "use_duration_discriminator": false,
+    "use_wavlm_discriminator": true,
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3, 7, 11],
+    "resblock_dilation_sizes": [
+      [1, 3, 5],
+      [1, 3, 5],
+      [1, 3, 5]
+    ],
+    "upsample_rates": [8, 8, 2, 2, 2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [16, 16, 8, 2, 2],
+    "n_layers_q": 3,
+    "use_spectral_norm": false,
+    "gin_channels": 512,
+    "slm": {
+      "model": "./slm/wavlm-base-plus",
+      "sr": 16000,
+      "hidden": 768,
+      "nlayers": 13,
+      "initial_channel": 64
+    }
+  },
+  "version": "2.0-JP-Extra",
+  "model_name": "jvnv-F1-jp"
+}

model_assets/jvnv-F1-jp/jvnv-F1-jp_e182_s16000.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd324b2dd04b4a3384e0dbf4a268fd8a9bbedcfe80608fbd2a3aaaa44e474abe
+size 251150980

model_assets/jvnv-F1-jp/style_vectors.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f959bb45ed0922efc31ff24e9147253814f42cb1d2d1e2bb10391a9df368489
+size 7296

model_assets/jvnv-F2-jp/config.json

@@ -0,0 +1,92 @@
+{
+  "train": {
+    "log_interval": 200,
+    "eval_interval": 1000,
+    "seed": 42,
+    "epochs": 300,
+    "learning_rate": 0.0001,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 4,
+    "bf16_run": false,
+    "fp16_run": false,
+    "lr_decay": 0.99996,
+    "segment_size": 16384,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0,
+    "c_commit": 100,
+    "skip_optimizer": true,
+    "freeze_ZH_bert": false,
+    "freeze_JP_bert": false,
+    "freeze_EN_bert": false,
+    "freeze_emo": false,
+    "freeze_style": false
+  },
+  "data": {
+    "use_jp_extra": true,
+    "training_files": "/content/drive/MyDrive/Style-Bert-VITS2/Data/jvnv-F2/train.list",
+    "validation_files": "/content/drive/MyDrive/Style-Bert-VITS2/Data/jvnv-F2/val.list",
+    "max_wav_value": 32768.0,
+    "sampling_rate": 44100,
+    "filter_length": 2048,
+    "hop_length": 512,
+    "win_length": 2048,
+    "n_mel_channels": 128,
+    "mel_fmin": 0.0,
+    "mel_fmax": null,
+    "add_blank": true,
+    "n_speakers": 1,
+    "cleaned_text": true,
+    "spk2id": {
+      "jvnv-F2-jp": 0
+    },
+    "num_styles": 7,
+    "style2id": {
+      "Neutral": 0,
+      "Angry": 1,
+      "Disgust": 2,
+      "Fear": 3,
+      "Happy": 4,
+      "Sad": 5,
+      "Surprise": 6
+    }
+  },
+  "model": {
+    "use_spk_conditioned_encoder": true,
+    "use_noise_scaled_mas": true,
+    "use_mel_posterior_encoder": false,
+    "use_duration_discriminator": false,
+    "use_wavlm_discriminator": true,
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3, 7, 11],
+    "resblock_dilation_sizes": [
+      [1, 3, 5],
+      [1, 3, 5],
+      [1, 3, 5]
+    ],
+    "upsample_rates": [8, 8, 2, 2, 2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [16, 16, 8, 2, 2],
+    "n_layers_q": 3,
+    "use_spectral_norm": false,
+    "gin_channels": 512,
+    "slm": {
+      "model": "./slm/wavlm-base-plus",
+      "sr": 16000,
+      "hidden": 768,
+      "nlayers": 13,
+      "initial_channel": 64
+    }
+  },
+  "version": "2.0-JP-Extra",
+  "model_name": "jvnv-F2-jp"
+}

model_assets/jvnv-F2-jp/jvnv-F2_e166_s20000.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6289a6f30bb9795744815b9da764a3c8198b18652d9fddef82fff1e14f0e784
+size 251150980

model_assets/jvnv-F2-jp/style_vectors.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:900f8cde3a336d12193fec7b7d8e6c5dc77b3a5d719a9be3f8598389cd88e643
+size 7296

model_assets/jvnv-M1-jp/config.json

@@ -0,0 +1,92 @@
+{
+  "train": {
+    "log_interval": 200,
+    "eval_interval": 1000,
+    "seed": 42,
+    "epochs": 300,
+    "learning_rate": 0.0001,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 4,
+    "bf16_run": true,
+    "fp16_run": false,
+    "lr_decay": 0.99996,
+    "segment_size": 16384,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0,
+    "c_commit": 100,
+    "skip_optimizer": true,
+    "freeze_ZH_bert": false,
+    "freeze_JP_bert": false,
+    "freeze_EN_bert": false,
+    "freeze_emo": false,
+    "freeze_style": false
+  },
+  "data": {
+    "use_jp_extra": true,
+    "training_files": "Data/jvnv-M1-jp/train.list",
+    "validation_files": "Data/jvnv-M1-jp/val.list",
+    "max_wav_value": 32768.0,
+    "sampling_rate": 44100,
+    "filter_length": 2048,
+    "hop_length": 512,
+    "win_length": 2048,
+    "n_mel_channels": 128,
+    "mel_fmin": 0.0,
+    "mel_fmax": null,
+    "add_blank": true,
+    "n_speakers": 1,
+    "cleaned_text": true,
+    "spk2id": {
+      "jvnv-M1-jp": 0
+    },
+    "num_styles": 7,
+    "style2id": {
+      "Neutral": 0,
+      "Angry": 1,
+      "Disgust": 2,
+      "Fear": 3,
+      "Happy": 4,
+      "Sad": 5,
+      "Surprise": 6
+    }
+  },
+  "model": {
+    "use_spk_conditioned_encoder": true,
+    "use_noise_scaled_mas": true,
+    "use_mel_posterior_encoder": false,
+    "use_duration_discriminator": false,
+    "use_wavlm_discriminator": true,
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3, 7, 11],
+    "resblock_dilation_sizes": [
+      [1, 3, 5],
+      [1, 3, 5],
+      [1, 3, 5]
+    ],
+    "upsample_rates": [8, 8, 2, 2, 2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [16, 16, 8, 2, 2],
+    "n_layers_q": 3,
+    "use_spectral_norm": false,
+    "gin_channels": 512,
+    "slm": {
+      "model": "./slm/wavlm-base-plus",
+      "sr": 16000,
+      "hidden": 768,
+      "nlayers": 13,
+      "initial_channel": 64
+    }
+  },
+  "version": "2.0-JP-Extra",
+  "model_name": "jvnv-M1-jp"
+}

model_assets/jvnv-M1-jp/jvnv-M1-jp_e158_s14000.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d86765f1fe08dbba74cd06283e96b6941b3f232329fabbba9c30e6edc27887a
+size 251150980

model_assets/jvnv-M1-jp/style_vectors.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a925435e8c1c9efc8fc8e90e690655ab9a7bae00a790892e13e936510d04f05
+size 7296

model_assets/jvnv-M2-jp/config.json

@@ -0,0 +1,92 @@
+{
+  "train": {
+    "log_interval": 200,
+    "eval_interval": 1000,
+    "seed": 42,
+    "epochs": 300,
+    "learning_rate": 0.0001,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 4,
+    "bf16_run": true,
+    "fp16_run": false,
+    "lr_decay": 0.99996,
+    "segment_size": 16384,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0,
+    "c_commit": 100,
+    "skip_optimizer": true,
+    "freeze_ZH_bert": false,
+    "freeze_JP_bert": false,
+    "freeze_EN_bert": false,
+    "freeze_emo": false,
+    "freeze_style": false
+  },
+  "data": {
+    "use_jp_extra": true,
+    "training_files": "Data/jvnv-M2-jp/train.list",
+    "validation_files": "Data/jvnv-M2-jp/val.list",
+    "max_wav_value": 32768.0,
+    "sampling_rate": 44100,
+    "filter_length": 2048,
+    "hop_length": 512,
+    "win_length": 2048,
+    "n_mel_channels": 128,
+    "mel_fmin": 0.0,
+    "mel_fmax": null,
+    "add_blank": true,
+    "n_speakers": 1,
+    "cleaned_text": true,
+    "spk2id": {
+      "jvnv-M2-jp": 0
+    },
+    "num_styles": 7,
+    "style2id": {
+      "Neutral": 0,
+      "Angry": 1,
+      "Disgust": 2,
+      "Fear": 3,
+      "Happy": 4,
+      "Sad": 5,
+      "Surprise": 6
+    }
+  },
+  "model": {
+    "use_spk_conditioned_encoder": true,
+    "use_noise_scaled_mas": true,
+    "use_mel_posterior_encoder": false,
+    "use_duration_discriminator": false,
+    "use_wavlm_discriminator": true,
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3, 7, 11],
+    "resblock_dilation_sizes": [
+      [1, 3, 5],
+      [1, 3, 5],
+      [1, 3, 5]
+    ],
+    "upsample_rates": [8, 8, 2, 2, 2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [16, 16, 8, 2, 2],
+    "n_layers_q": 3,
+    "use_spectral_norm": false,
+    "gin_channels": 512,
+    "slm": {
+      "model": "./slm/wavlm-base-plus",
+      "sr": 16000,
+      "hidden": 768,
+      "nlayers": 13,
+      "initial_channel": 64
+    }
+  },
+  "version": "2.0-JP-Extra",
+  "model_name": "jvnv-M2-jp"
+}

model_assets/jvnv-M2-jp/jvnv-M2-jp_e159_s17000.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8245f39438076d36a3befd8aefb15c38830cef326c1f7c9d9c8e64b647645402
+size 251150980

model_assets/jvnv-M2-jp/style_vectors.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c965bb63fa4a759d41a8a4a3649333125d6497ae8a705d81b7d5c5bd2854797c
+size 7296

models_jp_extra.py

@@ -0,0 +1,1071 @@
+import math
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+import commons
+import modules
+import attentions
+import monotonic_align
+
+from torch.nn import Conv1d, ConvTranspose1d, Conv2d
+from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
+
+from commons import init_weights, get_padding
+from text import symbols, num_tones, num_languages
+
+
+class DurationDiscriminator(nn.Module):  # vits2
+    def __init__(
+        self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0
+    ):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.gin_channels = gin_channels
+
+        self.drop = nn.Dropout(p_dropout)
+        self.conv_1 = nn.Conv1d(
+            in_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.norm_1 = modules.LayerNorm(filter_channels)
+        self.conv_2 = nn.Conv1d(
+            filter_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.norm_2 = modules.LayerNorm(filter_channels)
+        self.dur_proj = nn.Conv1d(1, filter_channels, 1)
+
+        self.LSTM = nn.LSTM(
+            2 * filter_channels, filter_channels, batch_first=True, bidirectional=True
+        )
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, in_channels, 1)
+
+        self.output_layer = nn.Sequential(
+            nn.Linear(2 * filter_channels, 1), nn.Sigmoid()
+        )
+
+    def forward_probability(self, x, dur):
+        dur = self.dur_proj(dur)
+        x = torch.cat([x, dur], dim=1)
+        x = x.transpose(1, 2)
+        x, _ = self.LSTM(x)
+        output_prob = self.output_layer(x)
+        return output_prob
+
+    def forward(self, x, x_mask, dur_r, dur_hat, g=None):
+        x = torch.detach(x)
+        if g is not None:
+            g = torch.detach(g)
+            x = x + self.cond(g)
+        x = self.conv_1(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_1(x)
+        x = self.drop(x)
+        x = self.conv_2(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_2(x)
+        x = self.drop(x)
+
+        output_probs = []
+        for dur in [dur_r, dur_hat]:
+            output_prob = self.forward_probability(x, dur)
+            output_probs.append(output_prob)
+
+        return output_probs
+
+
+class TransformerCouplingBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        n_flows=4,
+        gin_channels=0,
+        share_parameter=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+
+        self.wn = (
+            attentions.FFT(
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers,
+                kernel_size,
+                p_dropout,
+                isflow=True,
+                gin_channels=self.gin_channels,
+            )
+            if share_parameter
+            else None
+        )
+
+        for i in range(n_flows):
+            self.flows.append(
+                modules.TransformerCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    n_layers,
+                    n_heads,
+                    p_dropout,
+                    filter_channels,
+                    mean_only=True,
+                    wn_sharing_parameter=self.wn,
+                    gin_channels=self.gin_channels,
+                )
+            )
+            self.flows.append(modules.Flip())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+
+class StochasticDurationPredictor(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        filter_channels,
+        kernel_size,
+        p_dropout,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super().__init__()
+        filter_channels = in_channels  # it needs to be removed from future version.
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.log_flow = modules.Log()
+        self.flows = nn.ModuleList()
+        self.flows.append(modules.ElementwiseAffine(2))
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3)
+            )
+            self.flows.append(modules.Flip())
+
+        self.post_pre = nn.Conv1d(1, filter_channels, 1)
+        self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
+        self.post_convs = modules.DDSConv(
+            filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout
+        )
+        self.post_flows = nn.ModuleList()
+        self.post_flows.append(modules.ElementwiseAffine(2))
+        for i in range(4):
+            self.post_flows.append(
+                modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3)
+            )
+            self.post_flows.append(modules.Flip())
+
+        self.pre = nn.Conv1d(in_channels, filter_channels, 1)
+        self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
+        self.convs = modules.DDSConv(
+            filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout
+        )
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
+
+    def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
+        x = torch.detach(x)
+        x = self.pre(x)
+        if g is not None:
+            g = torch.detach(g)
+            x = x + self.cond(g)
+        x = self.convs(x, x_mask)
+        x = self.proj(x) * x_mask
+
+        if not reverse:
+            flows = self.flows
+            assert w is not None
+
+            logdet_tot_q = 0
+            h_w = self.post_pre(w)
+            h_w = self.post_convs(h_w, x_mask)
+            h_w = self.post_proj(h_w) * x_mask
+            e_q = (
+                torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype)
+                * x_mask
+            )
+            z_q = e_q
+            for flow in self.post_flows:
+                z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
+                logdet_tot_q += logdet_q
+            z_u, z1 = torch.split(z_q, [1, 1], 1)
+            u = torch.sigmoid(z_u) * x_mask
+            z0 = (w - u) * x_mask
+            logdet_tot_q += torch.sum(
+                (F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1, 2]
+            )
+            logq = (
+                torch.sum(-0.5 * (math.log(2 * math.pi) + (e_q**2)) * x_mask, [1, 2])
+                - logdet_tot_q
+            )
+
+            logdet_tot = 0
+            z0, logdet = self.log_flow(z0, x_mask)
+            logdet_tot += logdet
+            z = torch.cat([z0, z1], 1)
+            for flow in flows:
+                z, logdet = flow(z, x_mask, g=x, reverse=reverse)
+                logdet_tot = logdet_tot + logdet
+            nll = (
+                torch.sum(0.5 * (math.log(2 * math.pi) + (z**2)) * x_mask, [1, 2])
+                - logdet_tot
+            )
+            return nll + logq  # [b]
+        else:
+            flows = list(reversed(self.flows))
+            flows = flows[:-2] + [flows[-1]]  # remove a useless vflow
+            z = (
+                torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype)
+                * noise_scale
+            )
+            for flow in flows:
+                z = flow(z, x_mask, g=x, reverse=reverse)
+            z0, z1 = torch.split(z, [1, 1], 1)
+            logw = z0
+            return logw
+
+
+class DurationPredictor(nn.Module):
+    def __init__(
+        self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0
+    ):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.gin_channels = gin_channels
+
+        self.drop = nn.Dropout(p_dropout)
+        self.conv_1 = nn.Conv1d(
+            in_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.norm_1 = modules.LayerNorm(filter_channels)
+        self.conv_2 = nn.Conv1d(
+            filter_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.norm_2 = modules.LayerNorm(filter_channels)
+        self.proj = nn.Conv1d(filter_channels, 1, 1)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, in_channels, 1)
+
+    def forward(self, x, x_mask, g=None):
+        x = torch.detach(x)
+        if g is not None:
+            g = torch.detach(g)
+            x = x + self.cond(g)
+        x = self.conv_1(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_1(x)
+        x = self.drop(x)
+        x = self.conv_2(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_2(x)
+        x = self.drop(x)
+        x = self.proj(x * x_mask)
+        return x * x_mask
+
+
+class Bottleneck(nn.Sequential):
+    def __init__(self, in_dim, hidden_dim):
+        c_fc1 = nn.Linear(in_dim, hidden_dim, bias=False)
+        c_fc2 = nn.Linear(in_dim, hidden_dim, bias=False)
+        super().__init__(*[c_fc1, c_fc2])
+
+
+class Block(nn.Module):
+    def __init__(self, in_dim, hidden_dim) -> None:
+        super().__init__()
+        self.norm = nn.LayerNorm(in_dim)
+        self.mlp = MLP(in_dim, hidden_dim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = x + self.mlp(self.norm(x))
+        return x
+
+
+class MLP(nn.Module):
+    def __init__(self, in_dim, hidden_dim):
+        super().__init__()
+        self.c_fc1 = nn.Linear(in_dim, hidden_dim, bias=False)
+        self.c_fc2 = nn.Linear(in_dim, hidden_dim, bias=False)
+        self.c_proj = nn.Linear(hidden_dim, in_dim, bias=False)
+
+    def forward(self, x: torch.Tensor):
+        x = F.silu(self.c_fc1(x)) * self.c_fc2(x)
+        x = self.c_proj(x)
+        return x
+
+
+class TextEncoder(nn.Module):
+    def __init__(
+        self,
+        n_vocab,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.n_vocab = n_vocab
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.gin_channels = gin_channels
+        self.emb = nn.Embedding(len(symbols), hidden_channels)
+        nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)
+        self.tone_emb = nn.Embedding(num_tones, hidden_channels)
+        nn.init.normal_(self.tone_emb.weight, 0.0, hidden_channels**-0.5)
+        self.language_emb = nn.Embedding(num_languages, hidden_channels)
+        nn.init.normal_(self.language_emb.weight, 0.0, hidden_channels**-0.5)
+        self.bert_proj = nn.Conv1d(1024, hidden_channels, 1)
+
+        # Remove emo_vq since it's not working well.
+        self.style_proj = nn.Linear(256, hidden_channels)
+
+        self.encoder = attentions.Encoder(
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            gin_channels=self.gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, tone, language, bert, style_vec, g=None):
+        bert_emb = self.bert_proj(bert).transpose(1, 2)
+        style_emb = self.style_proj(style_vec.unsqueeze(1))
+        x = (
+            self.emb(x)
+            + self.tone_emb(tone)
+            + self.language_emb(language)
+            + bert_emb
+            + style_emb
+        ) * math.sqrt(
+            self.hidden_channels
+        )  # [b, t, h]
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+
+        x = self.encoder(x * x_mask, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return x, m, logs, x_mask
+
+
+class ResidualCouplingBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ResidualCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    dilation_rate,
+                    n_layers,
+                    gin_channels=gin_channels,
+                    mean_only=True,
+                )
+            )
+            self.flows.append(modules.Flip())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+
+class PosteriorEncoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+
+        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
+        self.enc = modules.WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            gin_channels=gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, g=None):
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.pre(x) * x_mask
+        x = self.enc(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
+        return z, m, logs, x_mask
+
+
+class Generator(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels=0,
+    ):
+        super(Generator, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+    def forward(self, x, g=None):
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        print("Removing weight norm...")
+        for layer in self.ups:
+            remove_weight_norm(layer)
+        for layer in self.resblocks:
+            layer.remove_weight_norm()
+
+
+class DiscriminatorP(torch.nn.Module):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        self.use_spectral_norm = use_spectral_norm
+        norm_f = weight_norm if use_spectral_norm is False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(
+                    Conv2d(
+                        1,
+                        32,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        32,
+                        128,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        128,
+                        512,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        512,
+                        1024,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        1024,
+                        1024,
+                        (kernel_size, 1),
+                        1,
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+            ]
+        )
+        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for layer in self.convs:
+            x = layer(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class DiscriminatorS(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = weight_norm if use_spectral_norm is False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(Conv1d(1, 16, 15, 1, padding=7)),
+                norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
+                norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
+                norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
+                norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
+                norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
+            ]
+        )
+        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        fmap = []
+
+        for layer in self.convs:
+            x = layer(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminator, self).__init__()
+        periods = [2, 3, 5, 7, 11]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class WavLMDiscriminator(nn.Module):
+    """docstring for Discriminator."""
+
+    def __init__(
+        self, slm_hidden=768, slm_layers=13, initial_channel=64, use_spectral_norm=False
+    ):
+        super(WavLMDiscriminator, self).__init__()
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.pre = norm_f(
+            Conv1d(slm_hidden * slm_layers, initial_channel, 1, 1, padding=0)
+        )
+
+        self.convs = nn.ModuleList(
+            [
+                norm_f(
+                    nn.Conv1d(
+                        initial_channel, initial_channel * 2, kernel_size=5, padding=2
+                    )
+                ),
+                norm_f(
+                    nn.Conv1d(
+                        initial_channel * 2,
+                        initial_channel * 4,
+                        kernel_size=5,
+                        padding=2,
+                    )
+                ),
+                norm_f(
+                    nn.Conv1d(initial_channel * 4, initial_channel * 4, 5, 1, padding=2)
+                ),
+            ]
+        )
+
+        self.conv_post = norm_f(Conv1d(initial_channel * 4, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        x = self.pre(x)
+
+        fmap = []
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x
+
+
+class ReferenceEncoder(nn.Module):
+    """
+    inputs --- [N, Ty/r, n_mels*r]  mels
+    outputs --- [N, ref_enc_gru_size]
+    """
+
+    def __init__(self, spec_channels, gin_channels=0):
+        super().__init__()
+        self.spec_channels = spec_channels
+        ref_enc_filters = [32, 32, 64, 64, 128, 128]
+        K = len(ref_enc_filters)
+        filters = [1] + ref_enc_filters
+        convs = [
+            weight_norm(
+                nn.Conv2d(
+                    in_channels=filters[i],
+                    out_channels=filters[i + 1],
+                    kernel_size=(3, 3),
+                    stride=(2, 2),
+                    padding=(1, 1),
+                )
+            )
+            for i in range(K)
+        ]
+        self.convs = nn.ModuleList(convs)
+        # self.wns = nn.ModuleList([weight_norm(num_features=ref_enc_filters[i]) for i in range(K)]) # noqa: E501
+
+        out_channels = self.calculate_channels(spec_channels, 3, 2, 1, K)
+        self.gru = nn.GRU(
+            input_size=ref_enc_filters[-1] * out_channels,
+            hidden_size=256 // 2,
+            batch_first=True,
+        )
+        self.proj = nn.Linear(128, gin_channels)
+
+    def forward(self, inputs, mask=None):
+        N = inputs.size(0)
+        out = inputs.view(N, 1, -1, self.spec_channels)  # [N, 1, Ty, n_freqs]
+        for conv in self.convs:
+            out = conv(out)
+            # out = wn(out)
+            out = F.relu(out)  # [N, 128, Ty//2^K, n_mels//2^K]
+
+        out = out.transpose(1, 2)  # [N, Ty//2^K, 128, n_mels//2^K]
+        T = out.size(1)
+        N = out.size(0)
+        out = out.contiguous().view(N, T, -1)  # [N, Ty//2^K, 128*n_mels//2^K]
+
+        self.gru.flatten_parameters()
+        memory, out = self.gru(out)  # out --- [1, N, 128]
+
+        return self.proj(out.squeeze(0))
+
+    def calculate_channels(self, L, kernel_size, stride, pad, n_convs):
+        for i in range(n_convs):
+            L = (L - kernel_size + 2 * pad) // stride + 1
+        return L
+
+
+class SynthesizerTrn(nn.Module):
+    """
+    Synthesizer for Training
+    """
+
+    def __init__(
+        self,
+        n_vocab,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        n_speakers=256,
+        gin_channels=256,
+        use_sdp=True,
+        n_flow_layer=4,
+        n_layers_trans_flow=6,
+        flow_share_parameter=False,
+        use_transformer_flow=True,
+        **kwargs
+    ):
+        super().__init__()
+        self.n_vocab = n_vocab
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.n_speakers = n_speakers
+        self.gin_channels = gin_channels
+        self.n_layers_trans_flow = n_layers_trans_flow
+        self.use_spk_conditioned_encoder = kwargs.get(
+            "use_spk_conditioned_encoder", True
+        )
+        self.use_sdp = use_sdp
+        self.use_noise_scaled_mas = kwargs.get("use_noise_scaled_mas", False)
+        self.mas_noise_scale_initial = kwargs.get("mas_noise_scale_initial", 0.01)
+        self.noise_scale_delta = kwargs.get("noise_scale_delta", 2e-6)
+        self.current_mas_noise_scale = self.mas_noise_scale_initial
+        if self.use_spk_conditioned_encoder and gin_channels > 0:
+            self.enc_gin_channels = gin_channels
+        self.enc_p = TextEncoder(
+            n_vocab,
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            gin_channels=self.enc_gin_channels,
+        )
+        self.dec = Generator(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+        )
+        self.enc_q = PosteriorEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        if use_transformer_flow:
+            self.flow = TransformerCouplingBlock(
+                inter_channels,
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers_trans_flow,
+                5,
+                p_dropout,
+                n_flow_layer,
+                gin_channels=gin_channels,
+                share_parameter=flow_share_parameter,
+            )
+        else:
+            self.flow = ResidualCouplingBlock(
+                inter_channels,
+                hidden_channels,
+                5,
+                1,
+                n_flow_layer,
+                gin_channels=gin_channels,
+            )
+        self.sdp = StochasticDurationPredictor(
+            hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels
+        )
+        self.dp = DurationPredictor(
+            hidden_channels, 256, 3, 0.5, gin_channels=gin_channels
+        )
+
+        if n_speakers >= 1:
+            self.emb_g = nn.Embedding(n_speakers, gin_channels)
+        else:
+            self.ref_enc = ReferenceEncoder(spec_channels, gin_channels)
+
+    def forward(
+        self,
+        x,
+        x_lengths,
+        y,
+        y_lengths,
+        sid,
+        tone,
+        language,
+        bert,
+        style_vec,
+    ):
+        if self.n_speakers > 0:
+            g = self.emb_g(sid).unsqueeze(-1)  # [b, h, 1]
+        else:
+            g = self.ref_enc(y.transpose(1, 2)).unsqueeze(-1)
+        x, m_p, logs_p, x_mask = self.enc_p(
+            x, x_lengths, tone, language, bert, style_vec, g=g
+        )
+        z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
+        z_p = self.flow(z, y_mask, g=g)
+
+        with torch.no_grad():
+            # negative cross-entropy
+            s_p_sq_r = torch.exp(-2 * logs_p)  # [b, d, t]
+            neg_cent1 = torch.sum(
+                -0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True
+            )  # [b, 1, t_s]
+            neg_cent2 = torch.matmul(
+                -0.5 * (z_p**2).transpose(1, 2), s_p_sq_r
+            )  # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]
+            neg_cent3 = torch.matmul(
+                z_p.transpose(1, 2), (m_p * s_p_sq_r)
+            )  # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]
+            neg_cent4 = torch.sum(
+                -0.5 * (m_p**2) * s_p_sq_r, [1], keepdim=True
+            )  # [b, 1, t_s]
+            neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4
+            if self.use_noise_scaled_mas:
+                epsilon = (
+                    torch.std(neg_cent)
+                    * torch.randn_like(neg_cent)
+                    * self.current_mas_noise_scale
+                )
+                neg_cent = neg_cent + epsilon
+
+            attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
+            attn = (
+                monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1))
+                .unsqueeze(1)
+                .detach()
+            )
+
+        w = attn.sum(2)
+
+        l_length_sdp = self.sdp(x, x_mask, w, g=g)
+        l_length_sdp = l_length_sdp / torch.sum(x_mask)
+
+        logw_ = torch.log(w + 1e-6) * x_mask
+        logw = self.dp(x, x_mask, g=g)
+        # logw_sdp = self.sdp(x, x_mask, g=g, reverse=True, noise_scale=1.0)
+        l_length_dp = torch.sum((logw - logw_) ** 2, [1, 2]) / torch.sum(
+            x_mask
+        )  # for averaging
+        # l_length_sdp += torch.sum((logw_sdp - logw_) ** 2, [1, 2]) / torch.sum(x_mask)
+
+        l_length = l_length_dp + l_length_sdp
+
+        # expand prior
+        m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)
+        logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)
+
+        z_slice, ids_slice = commons.rand_slice_segments(
+            z, y_lengths, self.segment_size
+        )
+        o = self.dec(z_slice, g=g)
+        return (
+            o,
+            l_length,
+            attn,
+            ids_slice,
+            x_mask,
+            y_mask,
+            (z, z_p, m_p, logs_p, m_q, logs_q),
+            (x, logw, logw_),  # , logw_sdp),
+            g,
+        )
+
+    def infer(
+        self,
+        x,
+        x_lengths,
+        sid,
+        tone,
+        language,
+        bert,
+        style_vec,
+        noise_scale=0.667,
+        length_scale=1,
+        noise_scale_w=0.8,
+        max_len=None,
+        sdp_ratio=0,
+        y=None,
+    ):
+        # x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths, tone, language, bert)
+        # g = self.gst(y)
+        if self.n_speakers > 0:
+            g = self.emb_g(sid).unsqueeze(-1)  # [b, h, 1]
+        else:
+            g = self.ref_enc(y.transpose(1, 2)).unsqueeze(-1)
+        x, m_p, logs_p, x_mask = self.enc_p(
+            x, x_lengths, tone, language, bert, style_vec, g=g
+        )
+        logw = self.sdp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w) * (
+            sdp_ratio
+        ) + self.dp(x, x_mask, g=g) * (1 - sdp_ratio)
+        w = torch.exp(logw) * x_mask * length_scale
+        w_ceil = torch.ceil(w)
+        y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
+        y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(
+            x_mask.dtype
+        )
+        attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
+        attn = commons.generate_path(w_ceil, attn_mask)
+
+        m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(
+            1, 2
+        )  # [b, t', t], [b, t, d] -> [b, d, t']
+        logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(
+            1, 2
+        )  # [b, t', t], [b, t, d] -> [b, d, t']
+
+        z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
+        z = self.flow(z_p, y_mask, g=g, reverse=True)
+        o = self.dec((z * y_mask)[:, :, :max_len], g=g)
+        return o, attn, y_mask, (z, z_p, m_p, logs_p)

monotonic_align/__init__.py

@@ -1,16 +1,16 @@
-from numpy import zeros, int32, float32
-from torch import from_numpy
-
-from .core import maximum_path_jit
-
-
-def maximum_path(neg_cent, mask):
-    device = neg_cent.device
-    dtype = neg_cent.dtype
-    neg_cent = neg_cent.data.cpu().numpy().astype(float32)
-    path = zeros(neg_cent.shape, dtype=int32)
-
-    t_t_max = mask.sum(1)[:, 0].data.cpu().numpy().astype(int32)
-    t_s_max = mask.sum(2)[:, 0].data.cpu().numpy().astype(int32)
-    maximum_path_jit(path, neg_cent, t_t_max, t_s_max)
-    return from_numpy(path).to(device=device, dtype=dtype)
+from numpy import zeros, int32, float32
+from torch import from_numpy
+
+from .core import maximum_path_jit
+
+
+def maximum_path(neg_cent, mask):
+    device = neg_cent.device
+    dtype = neg_cent.dtype
+    neg_cent = neg_cent.data.cpu().numpy().astype(float32)
+    path = zeros(neg_cent.shape, dtype=int32)
+
+    t_t_max = mask.sum(1)[:, 0].data.cpu().numpy().astype(int32)
+    t_s_max = mask.sum(2)[:, 0].data.cpu().numpy().astype(int32)
+    maximum_path_jit(path, neg_cent, t_t_max, t_s_max)
+    return from_numpy(path).to(device=device, dtype=dtype)

monotonic_align/core.py

@@ -1,46 +1,46 @@
-import numba
-
-
-@numba.jit(
-    numba.void(
-        numba.int32[:, :, ::1],
-        numba.float32[:, :, ::1],
-        numba.int32[::1],
-        numba.int32[::1],
-    ),
-    nopython=True,
-    nogil=True,
-)
-def maximum_path_jit(paths, values, t_ys, t_xs):
-    b = paths.shape[0]
-    max_neg_val = -1e9
-    for i in range(int(b)):
-        path = paths[i]
-        value = values[i]
-        t_y = t_ys[i]
-        t_x = t_xs[i]
-
-        v_prev = v_cur = 0.0
-        index = t_x - 1
-
-        for y in range(t_y):
-            for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
-                if x == y:
-                    v_cur = max_neg_val
-                else:
-                    v_cur = value[y - 1, x]
-                if x == 0:
-                    if y == 0:
-                        v_prev = 0.0
-                    else:
-                        v_prev = max_neg_val
-                else:
-                    v_prev = value[y - 1, x - 1]
-                value[y, x] += max(v_prev, v_cur)
-
-        for y in range(t_y - 1, -1, -1):
-            path[y, index] = 1
-            if index != 0 and (
-                index == y or value[y - 1, index] < value[y - 1, index - 1]
-            ):
-                index = index - 1
+import numba
+
+
+@numba.jit(
+    numba.void(
+        numba.int32[:, :, ::1],
+        numba.float32[:, :, ::1],
+        numba.int32[::1],
+        numba.int32[::1],
+    ),
+    nopython=True,
+    nogil=True,
+)
+def maximum_path_jit(paths, values, t_ys, t_xs):
+    b = paths.shape[0]
+    max_neg_val = -1e9
+    for i in range(int(b)):
+        path = paths[i]
+        value = values[i]
+        t_y = t_ys[i]
+        t_x = t_xs[i]
+
+        v_prev = v_cur = 0.0
+        index = t_x - 1
+
+        for y in range(t_y):
+            for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
+                if x == y:
+                    v_cur = max_neg_val
+                else:
+                    v_cur = value[y - 1, x]
+                if x == 0:
+                    if y == 0:
+                        v_prev = 0.0
+                    else:
+                        v_prev = max_neg_val
+                else:
+                    v_prev = value[y - 1, x - 1]
+                value[y, x] += max(v_prev, v_cur)
+
+        for y in range(t_y - 1, -1, -1):
+            path[y, index] = 1
+            if index != 0 and (
+                index == y or value[y - 1, index] < value[y - 1, index - 1]
+            ):
+                index = index - 1

requirements.txt

@@ -1,6 +1,5 @@
 cmudict
 cn2an
-faster-whisper>=0.10.0
 g2p_en
 GPUtil
 gradio
@@ -15,14 +14,12 @@ num2words
 numba
 numpy
 psutil
-pyannote.audio>=3.1.0
 pyopenjtalk-prebuilt
 pypinyin
 PyYAML
 requests
-sentencepiece
 safetensors
 scipy
 tensorboard
-torch
+torch>=2.1,<2.2 # For users without GPU or colab
 transformers

text/__init__.py

@@ -0,0 +1,32 @@
+from text.symbols import *
+
+_symbol_to_id = {s: i for i, s in enumerate(symbols)}
+
+
+def cleaned_text_to_sequence(cleaned_text, tones, language):
+    """Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
+    Args:
+      text: string to convert to a sequence
+    Returns:
+      List of integers corresponding to the symbols in the text
+    """
+    phones = [_symbol_to_id[symbol] for symbol in cleaned_text]
+    tone_start = language_tone_start_map[language]
+    tones = [i + tone_start for i in tones]
+    lang_id = language_id_map[language]
+    lang_ids = [lang_id for i in phones]
+    return phones, tones, lang_ids
+
+
+def get_bert(
+    norm_text, word2ph, language, device, assist_text=None, assist_text_weight=0.7
+):
+    from .chinese_bert import get_bert_feature as zh_bert
+    from .english_bert_mock import get_bert_feature as en_bert
+    from .japanese_bert import get_bert_feature as jp_bert
+
+    lang_bert_func_map = {"ZH": zh_bert, "EN": en_bert, "JP": jp_bert}
+    bert = lang_bert_func_map[language](
+        norm_text, word2ph, device, assist_text, assist_text_weight
+    )
+    return bert

text/chinese.py

@@ -0,0 +1,199 @@
+import os
+import re
+
+import cn2an
+from pypinyin import lazy_pinyin, Style
+
+from text.symbols import punctuation
+from text.tone_sandhi import ToneSandhi
+
+current_file_path = os.path.dirname(__file__)
+pinyin_to_symbol_map = {
+    line.split("\t")[0]: line.strip().split("\t")[1]
+    for line in open(os.path.join(current_file_path, "opencpop-strict.txt")).readlines()
+}
+
+import jieba.posseg as psg
+
+
+rep_map = {
+    "：": ",",
+    "；": ",",
+    "，": ",",
+    "。": ".",
+    "！": "!",
+    "？": "?",
+    "\n": ".",
+    "·": ",",
+    "、": ",",
+    "...": "…",
+    "$": ".",
+    "“": "'",
+    "”": "'",
+    '"': "'",
+    "‘": "'",
+    "’": "'",
+    "（": "'",
+    "）": "'",
+    "(": "'",
+    ")": "'",
+    "《": "'",
+    "》": "'",
+    "【": "'",
+    "】": "'",
+    "[": "'",
+    "]": "'",
+    "—": "-",
+    "～": "-",
+    "~": "-",
+    "「": "'",
+    "」": "'",
+}
+
+tone_modifier = ToneSandhi()
+
+
+def replace_punctuation(text):
+    text = text.replace("嗯", "恩").replace("呣", "母")
+    pattern = re.compile("|".join(re.escape(p) for p in rep_map.keys()))
+
+    replaced_text = pattern.sub(lambda x: rep_map[x.group()], text)
+
+    replaced_text = re.sub(
+        r"[^\u4e00-\u9fa5" + "".join(punctuation) + r"]+", "", replaced_text
+    )
+
+    return replaced_text
+
+
+def g2p(text):
+    pattern = r"(?<=[{0}])\s*".format("".join(punctuation))
+    sentences = [i for i in re.split(pattern, text) if i.strip() != ""]
+    phones, tones, word2ph = _g2p(sentences)
+    assert sum(word2ph) == len(phones)
+    assert len(word2ph) == len(text)  # Sometimes it will crash,you can add a try-catch.
+    phones = ["_"] + phones + ["_"]
+    tones = [0] + tones + [0]
+    word2ph = [1] + word2ph + [1]
+    return phones, tones, word2ph
+
+
+def _get_initials_finals(word):
+    initials = []
+    finals = []
+    orig_initials = lazy_pinyin(word, neutral_tone_with_five=True, style=Style.INITIALS)
+    orig_finals = lazy_pinyin(
+        word, neutral_tone_with_five=True, style=Style.FINALS_TONE3
+    )
+    for c, v in zip(orig_initials, orig_finals):
+        initials.append(c)
+        finals.append(v)
+    return initials, finals
+
+
+def _g2p(segments):
+    phones_list = []
+    tones_list = []
+    word2ph = []
+    for seg in segments:
+        # Replace all English words in the sentence
+        seg = re.sub("[a-zA-Z]+", "", seg)
+        seg_cut = psg.lcut(seg)
+        initials = []
+        finals = []
+        seg_cut = tone_modifier.pre_merge_for_modify(seg_cut)
+        for word, pos in seg_cut:
+            if pos == "eng":
+                continue
+            sub_initials, sub_finals = _get_initials_finals(word)
+            sub_finals = tone_modifier.modified_tone(word, pos, sub_finals)
+            initials.append(sub_initials)
+            finals.append(sub_finals)
+
+            # assert len(sub_initials) == len(sub_finals) == len(word)
+        initials = sum(initials, [])
+        finals = sum(finals, [])
+        #
+        for c, v in zip(initials, finals):
+            raw_pinyin = c + v
+            # NOTE: post process for pypinyin outputs
+            # we discriminate i, ii and iii
+            if c == v:
+                assert c in punctuation
+                phone = [c]
+                tone = "0"
+                word2ph.append(1)
+            else:
+                v_without_tone = v[:-1]
+                tone = v[-1]
+
+                pinyin = c + v_without_tone
+                assert tone in "12345"
+
+                if c:
+                    # 多音节
+                    v_rep_map = {
+                        "uei": "ui",
+                        "iou": "iu",
+                        "uen": "un",
+                    }
+                    if v_without_tone in v_rep_map.keys():
+                        pinyin = c + v_rep_map[v_without_tone]
+                else:
+                    # 单音节
+                    pinyin_rep_map = {
+                        "ing": "ying",
+                        "i": "yi",
+                        "in": "yin",
+                        "u": "wu",
+                    }
+                    if pinyin in pinyin_rep_map.keys():
+                        pinyin = pinyin_rep_map[pinyin]
+                    else:
+                        single_rep_map = {
+                            "v": "yu",
+                            "e": "e",
+                            "i": "y",
+                            "u": "w",
+                        }
+                        if pinyin[0] in single_rep_map.keys():
+                            pinyin = single_rep_map[pinyin[0]] + pinyin[1:]
+
+                assert pinyin in pinyin_to_symbol_map.keys(), (pinyin, seg, raw_pinyin)
+                phone = pinyin_to_symbol_map[pinyin].split(" ")
+                word2ph.append(len(phone))
+
+            phones_list += phone
+            tones_list += [int(tone)] * len(phone)
+    return phones_list, tones_list, word2ph
+
+
+def text_normalize(text):
+    numbers = re.findall(r"\d+(?:\.?\d+)?", text)
+    for number in numbers:
+        text = text.replace(number, cn2an.an2cn(number), 1)
+    text = replace_punctuation(text)
+    return text
+
+
+def get_bert_feature(text, word2ph):
+    from text import chinese_bert
+
+    return chinese_bert.get_bert_feature(text, word2ph)
+
+
+if __name__ == "__main__":
+    from text.chinese_bert import get_bert_feature
+
+    text = "啊！但是《原神》是由,米哈\游自主，  [研发]的一款全.新开放世界.冒险游戏"
+    text = text_normalize(text)
+    print(text)
+    phones, tones, word2ph = g2p(text)
+    bert = get_bert_feature(text, word2ph)
+
+    print(phones, tones, word2ph, bert.shape)
+
+
+# # 示例用法
+# text = "这是一个示例文本：,你好！这是一个测试...."
+# print(g2p_paddle(text))  # 输出: 这是一个示例文本你好这是一个测试

text/chinese_bert.py

@@ -0,0 +1,121 @@
+import sys
+
+import torch
+from transformers import AutoModelForMaskedLM, AutoTokenizer
+
+from config import config
+
+LOCAL_PATH = "./bert/chinese-roberta-wwm-ext-large"
+
+tokenizer = AutoTokenizer.from_pretrained(LOCAL_PATH)
+
+models = dict()
+
+
+def get_bert_feature(
+    text,
+    word2ph,
+    device=config.bert_gen_config.device,
+    assist_text=None,
+    assist_text_weight=0.7,
+):
+    if (
+        sys.platform == "darwin"
+        and torch.backends.mps.is_available()
+        and device == "cpu"
+    ):
+        device = "mps"
+    if not device:
+        device = "cuda"
+    if device == "cuda" and not torch.cuda.is_available():
+        device = "cpu"
+    if device not in models.keys():
+        models[device] = AutoModelForMaskedLM.from_pretrained(LOCAL_PATH).to(device)
+    with torch.no_grad():
+        inputs = tokenizer(text, return_tensors="pt")
+        for i in inputs:
+            inputs[i] = inputs[i].to(device)
+        res = models[device](**inputs, output_hidden_states=True)
+        res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
+        if assist_text:
+            style_inputs = tokenizer(assist_text, return_tensors="pt")
+            for i in style_inputs:
+                style_inputs[i] = style_inputs[i].to(device)
+            style_res = models[device](**style_inputs, output_hidden_states=True)
+            style_res = torch.cat(style_res["hidden_states"][-3:-2], -1)[0].cpu()
+            style_res_mean = style_res.mean(0)
+    assert len(word2ph) == len(text) + 2
+    word2phone = word2ph
+    phone_level_feature = []
+    for i in range(len(word2phone)):
+        if assist_text:
+            repeat_feature = (
+                res[i].repeat(word2phone[i], 1) * (1 - assist_text_weight)
+                + style_res_mean.repeat(word2phone[i], 1) * assist_text_weight
+            )
+        else:
+            repeat_feature = res[i].repeat(word2phone[i], 1)
+        phone_level_feature.append(repeat_feature)
+
+    phone_level_feature = torch.cat(phone_level_feature, dim=0)
+
+    return phone_level_feature.T
+
+
+if __name__ == "__main__":
+    word_level_feature = torch.rand(38, 1024)  # 12个词,每个词1024维特征
+    word2phone = [
+        1,
+        2,
+        1,
+        2,
+        2,
+        1,
+        2,
+        2,
+        1,
+        2,
+        2,
+        1,
+        2,
+        2,
+        2,
+        2,
+        2,
+        1,
+        1,
+        2,
+        2,
+        1,
+        2,
+        2,
+        2,
+        2,
+        1,
+        2,
+        2,
+        2,
+        2,
+        2,
+        1,
+        2,
+        2,
+        2,
+        2,
+        1,
+    ]
+
+    # 计算总帧数
+    total_frames = sum(word2phone)
+    print(word_level_feature.shape)
+    print(word2phone)
+    phone_level_feature = []
+    for i in range(len(word2phone)):
+        print(word_level_feature[i].shape)
+
+        # 对每个词重复word2phone[i]次
+        repeat_feature = word_level_feature[i].repeat(word2phone[i], 1)
+        phone_level_feature.append(repeat_feature)
+
+    phone_level_feature = torch.cat(phone_level_feature, dim=0)
+    print(phone_level_feature.shape)  # torch.Size([36, 1024])

text/cleaner.py

@@ -0,0 +1,31 @@
+from text import chinese, japanese, english, cleaned_text_to_sequence
+
+
+language_module_map = {"ZH": chinese, "JP": japanese, "EN": english}
+
+
+def clean_text(text, language, use_jp_extra=True):
+    language_module = language_module_map[language]
+    norm_text = language_module.text_normalize(text)
+    if language == "JP":
+        phones, tones, word2ph = language_module.g2p(norm_text, use_jp_extra)
+    else:
+        phones, tones, word2ph = language_module.g2p(norm_text)
+    return norm_text, phones, tones, word2ph
+
+
+def clean_text_bert(text, language):
+    language_module = language_module_map[language]
+    norm_text = language_module.text_normalize(text)
+    phones, tones, word2ph = language_module.g2p(norm_text)
+    bert = language_module.get_bert_feature(norm_text, word2ph)
+    return phones, tones, bert
+
+
+def text_to_sequence(text, language):
+    norm_text, phones, tones, word2ph = clean_text(text, language)
+    return cleaned_text_to_sequence(phones, tones, language)
+
+
+if __name__ == "__main__":
+    pass

text/cmudict_cache.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9b21b20325471934ba92f2e4a5976989e7d920caa32e7a286eacb027d197949
+size 6212655

text/english.py

@@ -0,0 +1,495 @@
+import pickle
+import os
+import re
+from g2p_en import G2p
+from transformers import DebertaV2Tokenizer
+
+from text import symbols
+from text.symbols import punctuation
+
+current_file_path = os.path.dirname(__file__)
+CMU_DICT_PATH = os.path.join(current_file_path, "cmudict.rep")
+CACHE_PATH = os.path.join(current_file_path, "cmudict_cache.pickle")
+_g2p = G2p()
+LOCAL_PATH = "./bert/deberta-v3-large"
+tokenizer = DebertaV2Tokenizer.from_pretrained(LOCAL_PATH)
+
+arpa = {
+    "AH0",
+    "S",
+    "AH1",
+    "EY2",
+    "AE2",
+    "EH0",
+    "OW2",
+    "UH0",
+    "NG",
+    "B",
+    "G",
+    "AY0",
+    "M",
+    "AA0",
+    "F",
+    "AO0",
+    "ER2",
+    "UH1",
+    "IY1",
+    "AH2",
+    "DH",
+    "IY0",
+    "EY1",
+    "IH0",
+    "K",
+    "N",
+    "W",
+    "IY2",
+    "T",
+    "AA1",
+    "ER1",
+    "EH2",
+    "OY0",
+    "UH2",
+    "UW1",
+    "Z",
+    "AW2",
+    "AW1",
+    "V",
+    "UW2",
+    "AA2",
+    "ER",
+    "AW0",
+    "UW0",
+    "R",
+    "OW1",
+    "EH1",
+    "ZH",
+    "AE0",
+    "IH2",
+    "IH",
+    "Y",
+    "JH",
+    "P",
+    "AY1",
+    "EY0",
+    "OY2",
+    "TH",
+    "HH",
+    "D",
+    "ER0",
+    "CH",
+    "AO1",
+    "AE1",
+    "AO2",
+    "OY1",
+    "AY2",
+    "IH1",
+    "OW0",
+    "L",
+    "SH",
+}
+
+
+def post_replace_ph(ph):
+    rep_map = {
+        "：": ",",
+        "；": ",",
+        "，": ",",
+        "。": ".",
+        "！": "!",
+        "？": "?",
+        "\n": ".",
+        "·": ",",
+        "、": ",",
+        "…": "...",
+        "···": "...",
+        "・・・": "...",
+        "v": "V",
+    }
+    if ph in rep_map.keys():
+        ph = rep_map[ph]
+    if ph in symbols:
+        return ph
+    if ph not in symbols:
+        ph = "UNK"
+    return ph
+
+
+rep_map = {
+    "：": ",",
+    "；": ",",
+    "，": ",",
+    "。": ".",
+    "！": "!",
+    "？": "?",
+    "\n": ".",
+    "．": ".",
+    "…": "...",
+    "···": "...",
+    "・・・": "...",
+    "·": ",",
+    "・": ",",
+    "、": ",",
+    "$": ".",
+    "“": "'",
+    "”": "'",
+    '"': "'",
+    "‘": "'",
+    "’": "'",
+    "（": "'",
+    "）": "'",
+    "(": "'",
+    ")": "'",
+    "《": "'",
+    "》": "'",
+    "【": "'",
+    "】": "'",
+    "[": "'",
+    "]": "'",
+    "—": "-",
+    "−": "-",
+    "～": "-",
+    "~": "-",
+    "「": "'",
+    "」": "'",
+}
+
+
+def replace_punctuation(text):
+    pattern = re.compile("|".join(re.escape(p) for p in rep_map.keys()))
+
+    replaced_text = pattern.sub(lambda x: rep_map[x.group()], text)
+
+    # replaced_text = re.sub(
+    #     r"[^\u3040-\u309F\u30A0-\u30FF\u4E00-\u9FFF\u3400-\u4DBF\u3005"
+    #     + "".join(punctuation)
+    #     + r"]+",
+    #     "",
+    #     replaced_text,
+    # )
+
+    return replaced_text
+
+
+def read_dict():
+    g2p_dict = {}
+    start_line = 49
+    with open(CMU_DICT_PATH) as f:
+        line = f.readline()
+        line_index = 1
+        while line:
+            if line_index >= start_line:
+                line = line.strip()
+                word_split = line.split("  ")
+                word = word_split[0]
+
+                syllable_split = word_split[1].split(" - ")
+                g2p_dict[word] = []
+                for syllable in syllable_split:
+                    phone_split = syllable.split(" ")
+                    g2p_dict[word].append(phone_split)
+
+            line_index = line_index + 1
+            line = f.readline()
+
+    return g2p_dict
+
+
+def cache_dict(g2p_dict, file_path):
+    with open(file_path, "wb") as pickle_file:
+        pickle.dump(g2p_dict, pickle_file)
+
+
+def get_dict():
+    if os.path.exists(CACHE_PATH):
+        with open(CACHE_PATH, "rb") as pickle_file:
+            g2p_dict = pickle.load(pickle_file)
+    else:
+        g2p_dict = read_dict()
+        cache_dict(g2p_dict, CACHE_PATH)
+
+    return g2p_dict
+
+
+eng_dict = get_dict()
+
+
+def refine_ph(phn):
+    tone = 0
+    if re.search(r"\d$", phn):
+        tone = int(phn[-1]) + 1
+        phn = phn[:-1]
+    else:
+        tone = 3
+    return phn.lower(), tone
+
+
+def refine_syllables(syllables):
+    tones = []
+    phonemes = []
+    for phn_list in syllables:
+        for i in range(len(phn_list)):
+            phn = phn_list[i]
+            phn, tone = refine_ph(phn)
+            phonemes.append(phn)
+            tones.append(tone)
+    return phonemes, tones
+
+
+import re
+import inflect
+
+_inflect = inflect.engine()
+_comma_number_re = re.compile(r"([0-9][0-9\,]+[0-9])")
+_decimal_number_re = re.compile(r"([0-9]+\.[0-9]+)")
+_pounds_re = re.compile(r"£([0-9\,]*[0-9]+)")
+_dollars_re = re.compile(r"\$([0-9\.\,]*[0-9]+)")
+_ordinal_re = re.compile(r"[0-9]+(st|nd|rd|th)")
+_number_re = re.compile(r"[0-9]+")
+
+# List of (regular expression, replacement) pairs for abbreviations:
+_abbreviations = [
+    (re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1])
+    for x in [
+        ("mrs", "misess"),
+        ("mr", "mister"),
+        ("dr", "doctor"),
+        ("st", "saint"),
+        ("co", "company"),
+        ("jr", "junior"),
+        ("maj", "major"),
+        ("gen", "general"),
+        ("drs", "doctors"),
+        ("rev", "reverend"),
+        ("lt", "lieutenant"),
+        ("hon", "honorable"),
+        ("sgt", "sergeant"),
+        ("capt", "captain"),
+        ("esq", "esquire"),
+        ("ltd", "limited"),
+        ("col", "colonel"),
+        ("ft", "fort"),
+    ]
+]
+
+
+# List of (ipa, lazy ipa) pairs:
+_lazy_ipa = [
+    (re.compile("%s" % x[0]), x[1])
+    for x in [
+        ("r", "ɹ"),
+        ("æ", "e"),
+        ("ɑ", "a"),
+        ("ɔ", "o"),
+        ("ð", "z"),
+        ("θ", "s"),
+        ("ɛ", "e"),
+        ("ɪ", "i"),
+        ("ʊ", "u"),
+        ("ʒ", "ʥ"),
+        ("ʤ", "ʥ"),
+        ("ˈ", "↓"),
+    ]
+]
+
+# List of (ipa, lazy ipa2) pairs:
+_lazy_ipa2 = [
+    (re.compile("%s" % x[0]), x[1])
+    for x in [
+        ("r", "ɹ"),
+        ("ð", "z"),
+        ("θ", "s"),
+        ("ʒ", "ʑ"),
+        ("ʤ", "dʑ"),
+        ("ˈ", "↓"),
+    ]
+]
+
+# List of (ipa, ipa2) pairs
+_ipa_to_ipa2 = [
+    (re.compile("%s" % x[0]), x[1]) for x in [("r", "ɹ"), ("ʤ", "dʒ"), ("ʧ", "tʃ")]
+]
+
+
+def _expand_dollars(m):
+    match = m.group(1)
+    parts = match.split(".")
+    if len(parts) > 2:
+        return match + " dollars"  # Unexpected format
+    dollars = int(parts[0]) if parts[0] else 0
+    cents = int(parts[1]) if len(parts) > 1 and parts[1] else 0
+    if dollars and cents:
+        dollar_unit = "dollar" if dollars == 1 else "dollars"
+        cent_unit = "cent" if cents == 1 else "cents"
+        return "%s %s, %s %s" % (dollars, dollar_unit, cents, cent_unit)
+    elif dollars:
+        dollar_unit = "dollar" if dollars == 1 else "dollars"
+        return "%s %s" % (dollars, dollar_unit)
+    elif cents:
+        cent_unit = "cent" if cents == 1 else "cents"
+        return "%s %s" % (cents, cent_unit)
+    else:
+        return "zero dollars"
+
+
+def _remove_commas(m):
+    return m.group(1).replace(",", "")
+
+
+def _expand_ordinal(m):
+    return _inflect.number_to_words(m.group(0))
+
+
+def _expand_number(m):
+    num = int(m.group(0))
+    if num > 1000 and num < 3000:
+        if num == 2000:
+            return "two thousand"
+        elif num > 2000 and num < 2010:
+            return "two thousand " + _inflect.number_to_words(num % 100)
+        elif num % 100 == 0:
+            return _inflect.number_to_words(num // 100) + " hundred"
+        else:
+            return _inflect.number_to_words(
+                num, andword="", zero="oh", group=2
+            ).replace(", ", " ")
+    else:
+        return _inflect.number_to_words(num, andword="")
+
+
+def _expand_decimal_point(m):
+    return m.group(1).replace(".", " point ")
+
+
+def normalize_numbers(text):
+    text = re.sub(_comma_number_re, _remove_commas, text)
+    text = re.sub(_pounds_re, r"\1 pounds", text)
+    text = re.sub(_dollars_re, _expand_dollars, text)
+    text = re.sub(_decimal_number_re, _expand_decimal_point, text)
+    text = re.sub(_ordinal_re, _expand_ordinal, text)
+    text = re.sub(_number_re, _expand_number, text)
+    return text
+
+
+def text_normalize(text):
+    text = normalize_numbers(text)
+    text = replace_punctuation(text)
+    text = re.sub(r"([,;.\?\!])([\w])", r"\1 \2", text)
+    return text
+
+
+def distribute_phone(n_phone, n_word):
+    phones_per_word = [0] * n_word
+    for task in range(n_phone):
+        min_tasks = min(phones_per_word)
+        min_index = phones_per_word.index(min_tasks)
+        phones_per_word[min_index] += 1
+    return phones_per_word
+
+
+def sep_text(text):
+    words = re.split(r"([,;.\?\!\s+])", text)
+    words = [word for word in words if word.strip() != ""]
+    return words
+
+
+def text_to_words(text):
+    tokens = tokenizer.tokenize(text)
+    words = []
+    for idx, t in enumerate(tokens):
+        if t.startswith("▁"):
+            words.append([t[1:]])
+        else:
+            if t in punctuation:
+                if idx == len(tokens) - 1:
+                    words.append([f"{t}"])
+                else:
+                    if (
+                        not tokens[idx + 1].startswith("▁")
+                        and tokens[idx + 1] not in punctuation
+                    ):
+                        if idx == 0:
+                            words.append([])
+                        words[-1].append(f"{t}")
+                    else:
+                        words.append([f"{t}"])
+            else:
+                if idx == 0:
+                    words.append([])
+                words[-1].append(f"{t}")
+    return words
+
+
+def g2p(text):
+    phones = []
+    tones = []
+    phone_len = []
+    # words = sep_text(text)
+    # tokens = [tokenizer.tokenize(i) for i in words]
+    words = text_to_words(text)
+
+    for word in words:
+        temp_phones, temp_tones = [], []
+        if len(word) > 1:
+            if "'" in word:
+                word = ["".join(word)]
+        for w in word:
+            if w in punctuation:
+                temp_phones.append(w)
+                temp_tones.append(0)
+                continue
+            if w.upper() in eng_dict:
+                phns, tns = refine_syllables(eng_dict[w.upper()])
+                temp_phones += [post_replace_ph(i) for i in phns]
+                temp_tones += tns
+                # w2ph.append(len(phns))
+            else:
+                phone_list = list(filter(lambda p: p != " ", _g2p(w)))
+                phns = []
+                tns = []
+                for ph in phone_list:
+                    if ph in arpa:
+                        ph, tn = refine_ph(ph)
+                        phns.append(ph)
+                        tns.append(tn)
+                    else:
+                        phns.append(ph)
+                        tns.append(0)
+                temp_phones += [post_replace_ph(i) for i in phns]
+                temp_tones += tns
+        phones += temp_phones
+        tones += temp_tones
+        phone_len.append(len(temp_phones))
+        # phones = [post_replace_ph(i) for i in phones]
+
+    word2ph = []
+    for token, pl in zip(words, phone_len):
+        word_len = len(token)
+
+        aaa = distribute_phone(pl, word_len)
+        word2ph += aaa
+
+    phones = ["_"] + phones + ["_"]
+    tones = [0] + tones + [0]
+    word2ph = [1] + word2ph + [1]
+    assert len(phones) == len(tones), text
+    assert len(phones) == sum(word2ph), text
+
+    return phones, tones, word2ph
+
+
+def get_bert_feature(text, word2ph):
+    from text import english_bert_mock
+
+    return english_bert_mock.get_bert_feature(text, word2ph)
+
+
+if __name__ == "__main__":
+    # print(get_dict())
+    # print(eng_word_to_phoneme("hello"))
+    print(g2p("In this paper, we propose 1 DSPGAN, a GAN-based universal vocoder."))
+    # all_phones = set()
+    # for k, syllables in eng_dict.items():
+    #     for group in syllables:
+    #         for ph in group:
+    #             all_phones.add(ph)
+    # print(all_phones)

text/english_bert_mock.py

@@ -0,0 +1,63 @@
+import sys
+
+import torch
+from transformers import DebertaV2Model, DebertaV2Tokenizer
+
+from config import config
+
+
+LOCAL_PATH = "./bert/deberta-v3-large"
+
+tokenizer = DebertaV2Tokenizer.from_pretrained(LOCAL_PATH)
+
+models = dict()
+
+
+def get_bert_feature(
+    text,
+    word2ph,
+    device=config.bert_gen_config.device,
+    assist_text=None,
+    assist_text_weight=0.7,
+):
+    if (
+        sys.platform == "darwin"
+        and torch.backends.mps.is_available()
+        and device == "cpu"
+    ):
+        device = "mps"
+    if not device:
+        device = "cuda"
+    if device == "cuda" and not torch.cuda.is_available():
+        device = "cpu"
+    if device not in models.keys():
+        models[device] = DebertaV2Model.from_pretrained(LOCAL_PATH).to(device)
+    with torch.no_grad():
+        inputs = tokenizer(text, return_tensors="pt")
+        for i in inputs:
+            inputs[i] = inputs[i].to(device)
+        res = models[device](**inputs, output_hidden_states=True)
+        res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
+        if assist_text:
+            style_inputs = tokenizer(assist_text, return_tensors="pt")
+            for i in style_inputs:
+                style_inputs[i] = style_inputs[i].to(device)
+            style_res = models[device](**style_inputs, output_hidden_states=True)
+            style_res = torch.cat(style_res["hidden_states"][-3:-2], -1)[0].cpu()
+            style_res_mean = style_res.mean(0)
+    assert len(word2ph) == res.shape[0], (text, res.shape[0], len(word2ph))
+    word2phone = word2ph
+    phone_level_feature = []
+    for i in range(len(word2phone)):
+        if assist_text:
+            repeat_feature = (
+                res[i].repeat(word2phone[i], 1) * (1 - assist_text_weight)
+                + style_res_mean.repeat(word2phone[i], 1) * assist_text_weight
+            )
+        else:
+            repeat_feature = res[i].repeat(word2phone[i], 1)
+        phone_level_feature.append(repeat_feature)
+
+    phone_level_feature = torch.cat(phone_level_feature, dim=0)
+
+    return phone_level_feature.T

text/japanese.py

@@ -0,0 +1,585 @@
+# Convert Japanese text to phonemes which is
+# compatible with Julius https://github.com/julius-speech/segmentation-kit
+import re
+import unicodedata
+
+import pyopenjtalk
+from num2words import num2words
+from transformers import AutoTokenizer
+
+from common.log import logger
+from text import punctuation
+from text.japanese_mora_list import (
+    mora_kata_to_mora_phonemes,
+    mora_phonemes_to_mora_kata,
+)
+
+# 子音の集合
+COSONANTS = set(
+    [
+        cosonant
+        for cosonant, _ in mora_kata_to_mora_phonemes.values()
+        if cosonant is not None
+    ]
+)
+
+# 母音の集合、便宜上「ん」を含める
+VOWELS = {"a", "i", "u", "e", "o", "N"}
+
+
+# 正規化で記号を変換するための辞書
+rep_map = {
+    "：": ",",
+    "；": ",",
+    "，": ",",
+    "。": ".",
+    "！": "!",
+    "？": "?",
+    "\n": ".",
+    "．": ".",
+    "…": "...",
+    "···": "...",
+    "・・・": "...",
+    "·": ",",
+    "・": ",",
+    "、": ",",
+    "$": ".",
+    "“": "'",
+    "”": "'",
+    '"': "'",
+    "‘": "'",
+    "’": "'",
+    "（": "'",
+    "）": "'",
+    "(": "'",
+    ")": "'",
+    "《": "'",
+    "》": "'",
+    "【": "'",
+    "】": "'",
+    "[": "'",
+    "]": "'",
+    "—": "-",
+    "−": "-",
+    # "～": "-",  # これは長音記号「ー」として扱うよう変更
+    # "~": "-",  # これも長音記号「ー」として扱うよう変更
+    "「": "'",
+    "」": "'",
+}
+
+
+def text_normalize(text):
+    """
+    日本語のテキストを正規化する。
+    結果は、ちょうど次の文字のみからなる：
+    - ひらがな
+    - カタカナ（全角長音記号「ー」が入る！）
+    - 漢字
+    - 半角アルファベット（大文字と小文字）
+    - ギリシャ文字
+    - `.` （句点`。`や`…`の一部や改行等）
+    - `,` （読点`、`や`:`等）
+    - `?` （疑問符`？`）
+    - `!` （感嘆符`！`）
+    - `'` （`「`や`」`等）
+    - `-` （`―`（ダッシュ、長音記号ではない）や`-`等）
+
+    注意点:
+    - 三点リーダー`…`は`...`に変換される（`なるほど…。` → `なるほど....`）
+    - 数字は漢字に変換される（`1,100円` → `千百円`、`52.34` → `五十二点三四`）
+    - 読点や疑問符等の位置・個数等は保持される（`??あ、、！！！` → `??あ,,!!!`）
+    """
+    res = unicodedata.normalize("NFKC", text)  # ここでアルファベットは半角になる
+    res = japanese_convert_numbers_to_words(res)  # 「100円」→「百円」等
+    # 「～」と「~」も長音記号として扱う
+    res = res.replace("~", "ー")
+    res = res.replace("～", "ー")
+
+    res = replace_punctuation(res)  # 句読点等正規化、読めない文字を削除
+
+    # 結合文字の濁点・半濁点を削除
+    # 通常の「ば」等はそのままのこされる、「あ゛」は上で「あ゙」になりここで「あ」になる
+    res = res.replace("\u3099", "")  # 結合文字の濁点を削除、る゙ → る
+    res = res.replace("\u309A", "")  # 結合文字の半濁点を削除、な゚ → な
+    return res
+
+
+def replace_punctuation(text: str) -> str:
+    """句読点等を「.」「,」「!」「?」「'」「-」に正規化し、OpenJTalkで読みが取得できるもののみ残す：
+    漢字・平仮名・カタカナ、アルファベット、ギリシャ文字
+    """
+    pattern = re.compile("|".join(re.escape(p) for p in rep_map.keys()))
+
+    # 句読点を辞書で置換
+    replaced_text = pattern.sub(lambda x: rep_map[x.group()], text)
+
+    replaced_text = re.sub(
+        # ↓ ひらがな、カタカナ、漢字
+        r"[^\u3040-\u309F\u30A0-\u30FF\u4E00-\u9FFF\u3400-\u4DBF\u3005"
+        # ↓ 半角アルファベット（大文字と小文字）
+        + r"\u0041-\u005A\u0061-\u007A"
+        # ↓ 全角アルファベット（大文字と小文字）
+        + r"\uFF21-\uFF3A\uFF41-\uFF5A"
+        # ↓ ギリシャ文字
+        + r"\u0370-\u03FF\u1F00-\u1FFF"
+        # ↓ "!", "?", "…", ",", ".", "'", "-", 但し`…`はすでに`...`に変換されている
+        + "".join(punctuation) + r"]+",
+        # 上述以外の文字を削除
+        "",
+        replaced_text,
+    )
+
+    return replaced_text
+
+
+_NUMBER_WITH_SEPARATOR_RX = re.compile("[0-9]{1,3}(,[0-9]{3})+")
+_CURRENCY_MAP = {"$": "ドル", "¥": "円", "£": "ポンド", "€": "ユーロ"}
+_CURRENCY_RX = re.compile(r"([$¥£€])([0-9.]*[0-9])")
+_NUMBER_RX = re.compile(r"[0-9]+(\.[0-9]+)?")
+
+
+def japanese_convert_numbers_to_words(text: str) -> str:
+    res = _NUMBER_WITH_SEPARATOR_RX.sub(lambda m: m[0].replace(",", ""), text)
+    res = _CURRENCY_RX.sub(lambda m: m[2] + _CURRENCY_MAP.get(m[1], m[1]), res)
+    res = _NUMBER_RX.sub(lambda m: num2words(m[0], lang="ja"), res)
+    return res
+
+
+def g2p(
+    norm_text: str, use_jp_extra: bool = True
+) -> tuple[list[str], list[int], list[int]]:
+    """
+    他で使われるメインの関数。`text_normalize()`で正規化された`norm_text`を受け取り、
+    - phones: 音素のリスト（ただし`!`や`,`や`.`等punctuationが含まれうる）
+    - tones: アクセントのリスト、0（低）と1（高）からなり、phonesと同じ長さ
+    - word2ph: 元のテキストの各文字に音素が何個割り当てられるかを表すリスト
+    のタプルを返す。
+    ただし`phones`と`tones`の最初と終わりに`_`が入り、応じて`word2ph`の最初と最後に1が追加される。
+    use_jp_extra: Falseの場合、「ん」の音素を「N」ではなく「n」とする。
+    """
+    # pyopenjtalkのフルコンテキストラベルを使ってアクセントを取り出すと、punctuationの位置が消えてしまい情報が失われてしまう：
+    # 「こんにちは、世界。」と「こんにちは！世界。」と「こんにちは！！！？？？世界……。」は全て同じになる。
+    # よって、まずpunctuation無しの音素とアクセントのリストを作り、
+    # それとは別にpyopenjtalk.run_frontend()で得られる音素リスト（こちらはpunctuationが保持される）を使い、
+    # アクセント割当をしなおすことによってpunctuationを含めた音素とアクセントのリストを作る。
+
+    # punctuationがすべて消えた、音素とアクセントのタプルのリスト（「ん」は「N」）
+    phone_tone_list_wo_punct = g2phone_tone_wo_punct(norm_text)
+
+    # sep_text: 単語単位の単語のリスト
+    # sep_kata: 単語単位の単語のカタカナ読みのリスト
+    sep_text, sep_kata = text2sep_kata(norm_text)
+
+    # sep_phonemes: 各単語ごとの音素のリストのリスト
+    sep_phonemes = handle_long([kata2phoneme_list(i) for i in sep_kata])
+
+    # phone_w_punct: sep_phonemesを結合した、punctuationを元のまま保持した音素列
+    phone_w_punct: list[str] = []
+    for i in sep_phonemes:
+        phone_w_punct += i
+
+    # punctuation無しのアクセント情報を使って、punctuationを含めたアクセント情報を作る
+    phone_tone_list = align_tones(phone_w_punct, phone_tone_list_wo_punct)
+    # logger.debug(f"phone_tone_list:\n{phone_tone_list}")
+    # word2phは厳密な解答は不可能なので（「今日」「眼鏡」等の熟字訓が存在）、
+    # Bert-VITS2では、単語単位の分割を使って、単語の文字ごとにだいたい均等に音素を分配する
+
+    # sep_textから、各単語を1文字1文字分割して、文字のリスト（のリスト）を作る
+    sep_tokenized: list[list[str]] = []
+    for i in sep_text:
+        if i not in punctuation:
+            sep_tokenized.append(
+                tokenizer.tokenize(i)
+            )  # ここでおそらく`i`が文字単位に分割される
+        else:
+            sep_tokenized.append([i])
+
+    # 各単語について、音素の数と文字の数を比較して、均等っぽく分配する
+    word2ph = []
+    for token, phoneme in zip(sep_tokenized, sep_phonemes):
+        phone_len = len(phoneme)
+        word_len = len(token)
+        word2ph += distribute_phone(phone_len, word_len)
+
+    # 最初と最後に`_`記号を追加、アクセントは0（低）、word2phもそれに合わせて追加
+    phone_tone_list = [("_", 0)] + phone_tone_list + [("_", 0)]
+    word2ph = [1] + word2ph + [1]
+
+    phones = [phone for phone, _ in phone_tone_list]
+    tones = [tone for _, tone in phone_tone_list]
+
+    assert len(phones) == sum(word2ph), f"{len(phones)} != {sum(word2ph)}"
+
+    # use_jp_extraでない場合は「N」を「n」に変換
+    if not use_jp_extra:
+        phones = [phone if phone != "N" else "n" for phone in phones]
+
+    return phones, tones, word2ph
+
+
+def g2kata_tone(norm_text: str) -> list[tuple[str, int]]:
+    phones, tones, _ = g2p(norm_text, use_jp_extra=True)
+    return phone_tone2kata_tone(list(zip(phones, tones)))
+
+
+def phone_tone2kata_tone(phone_tone: list[tuple[str, int]]) -> list[tuple[str, int]]:
+    """phone_toneをのphone部分をカタカナに変換する。ただし最初と最後の("_", 0)は無視"""
+    phone_tone = phone_tone[1:]  # 最初の("_", 0)を無視
+    phones = [phone for phone, _ in phone_tone]
+    tones = [tone for _, tone in phone_tone]
+    result: list[tuple[str, int]] = []
+    current_mora = ""
+    for phone, next_phone, tone, next_tone in zip(phones, phones[1:], tones, tones[1:]):
+        # zipの関係で最後の("_", 0)は無視されている
+        if phone in punctuation:
+            result.append((phone, tone))
+            continue
+        if phone in COSONANTS:  # n以外の子音の場合
+            assert current_mora == "", f"Unexpected {phone} after {current_mora}"
+            assert tone == next_tone, f"Unexpected {phone} tone {tone} != {next_tone}"
+            current_mora = phone
+        else:
+            # phoneが母音もしくは「N」
+            current_mora += phone
+            result.append((mora_phonemes_to_mora_kata[current_mora], tone))
+            current_mora = ""
+    return result
+
+
+def kata_tone2phone_tone(kata_tone: list[tuple[str, int]]) -> list[tuple[str, int]]:
+    """`phone_tone2kata_tone()`の逆。"""
+    result: list[tuple[str, int]] = [("_", 0)]
+    for mora, tone in kata_tone:
+        if mora in punctuation:
+            result.append((mora, tone))
+        else:
+            cosonant, vowel = mora_kata_to_mora_phonemes[mora]
+            if cosonant is None:
+                result.append((vowel, tone))
+            else:
+                result.append((cosonant, tone))
+                result.append((vowel, tone))
+    result.append(("_", 0))
+    return result
+
+
+def g2phone_tone_wo_punct(text: str) -> list[tuple[str, int]]:
+    """
+    テキストに対して、音素とアクセント（0か1）のペアのリストを返す。
+    ただし「!」「.」「?」等の非音素記号(punctuation)は全て消える（ポーズ記号も残さない）。
+    非音素記号を含める処理は`align_tones()`で行われる。
+    また「っ」は「q」に、「ん」は「N」に変換される。
+    例: "こんにちは、世界ー。。元気？！" →
+    [('k', 0), ('o', 0), ('N', 1), ('n', 1), ('i', 1), ('ch', 1), ('i', 1), ('w', 1), ('a', 1), ('s', 1), ('e', 1), ('k', 0), ('a', 0), ('i', 0), ('i', 0), ('g', 1), ('e', 1), ('N', 0), ('k', 0), ('i', 0)]
+    """
+    prosodies = pyopenjtalk_g2p_prosody(text, drop_unvoiced_vowels=True)
+    # logger.debug(f"prosodies: {prosodies}")
+    result: list[tuple[str, int]] = []
+    current_phrase: list[tuple[str, int]] = []
+    current_tone = 0
+    for i, letter in enumerate(prosodies):
+        # 特殊記号の処理
+
+        # 文頭記号、無視する
+        if letter == "^":
+            assert i == 0, "Unexpected ^"
+        # アクセント句の終わりに来る記号
+        elif letter in ("$", "?", "_", "#"):
+            # 保持しているフレーズを、アクセント数値を0-1に修正し結果に追加
+            result.extend(fix_phone_tone(current_phrase))
+            # 末尾に来る終了記号、無視（文中の疑問文は`_`になる）
+            if letter in ("$", "?"):
+                assert i == len(prosodies) - 1, f"Unexpected {letter}"
+            # あとは"_"（ポーズ）と"#"（アクセント句の境界）のみ
+            # これらは残さず、次のアクセント句に備える。
+            current_phrase = []
+            # 0を基準点にしてそこから上昇・下降する（負の場合は上の`fix_phone_tone`で直る）
+            current_tone = 0
+        # アクセント上昇記号
+        elif letter == "[":
+            current_tone = current_tone + 1
+        # アクセント下降記号
+        elif letter == "]":
+            current_tone = current_tone - 1
+        # それ以外は通常の音素
+        else:
+            if letter == "cl":  # 「っ」の処理
+                letter = "q"
+            # elif letter == "N":  # 「ん」の処理
+            #     letter = "n"
+            current_phrase.append((letter, current_tone))
+    return result
+
+
+def text2sep_kata(norm_text: str) -> tuple[list[str], list[str]]:
+    """
+    `text_normalize`で正規化済みの`norm_text`を受け取り、それを単語分割し、
+    分割された単語リストとその読み（カタカナor記号1文字）のリストのタプルを返す。
+    単語分割結果は、`g2p()`の`word2ph`で1文字あたりに割り振る音素記号の数を決めるために使う。
+    例:
+    `私はそう思う!って感じ?` →
+    ["私", "は", "そう", "思う", "!", "って", "感じ", "?"], ["ワタシ", "ワ", "ソー", "オモウ", "!", "ッテ", "カンジ", "?"]
+    """
+    # parsed: OpenJTalkの解析結果
+    parsed = pyopenjtalk.run_frontend(norm_text)
+    sep_text: list[str] = []
+    sep_kata: list[str] = []
+    for parts in parsed:
+        # word: 実際の単語の文字列
+        # yomi: その読み、但し無声化サインの`’`は除去
+        word, yomi = replace_punctuation(parts["string"]), parts["pron"].replace(
+            "’", ""
+        )
+        """
+        ここで`yomi`の取りうる値は以下の通りのはず。
+        - `word`が通常単語 → 通常の読み（カタカナ）
+            （カタカナからなり、長音記号も含みうる、`アー` 等）
+        - `word`が`ー` から始まる → `ーラー` や `ーーー` など
+        - `word`が句読点や空白等 → `、`
+        - `word`が`?` → `？`（全角になる）
+        他にも`word`が読めないキリル文字アラビア文字等が来ると`、`になるが、正規化でこの場合は起きないはず。
+        また元のコードでは`yomi`が空白の場合の処理があったが、これは起きないはず。
+        処理すべきは`yomi`が`、`の場合のみのはず。
+        """
+        assert yomi != "", f"Empty yomi: {word}"
+        if yomi == "、":
+            # wordは正規化されているので、`.`, `,`, `!`, `'`, `-`, `--` のいずれか
+            if word not in (
+                ".",
+                ",",
+                "!",
+                "'",
+                "-",
+                "--",
+            ):
+                # ここはpyopenjtalkが読めない文字等のときに起こる
+                raise ValueError(f"Cannot read: {word} in:\n{norm_text}")
+            # yomiは元の記号のままに変更
+            yomi = word
+        elif yomi == "？":
+            assert word == "?", f"yomi `？` comes from: {word}"
+            yomi = "?"
+        sep_text.append(word)
+        sep_kata.append(yomi)
+    return sep_text, sep_kata
+
+
+# ESPnetの実装から引用、変更点無し。「ん」は「N」なことに注意。
+# https://github.com/espnet/espnet/blob/master/espnet2/text/phoneme_tokenizer.py
+def pyopenjtalk_g2p_prosody(text: str, drop_unvoiced_vowels: bool = True) -> list[str]:
+    """Extract phoneme + prosoody symbol sequence from input full-context labels.
+
+    The algorithm is based on `Prosodic features control by symbols as input of
+    sequence-to-sequence acoustic modeling for neural TTS`_ with some r9y9's tweaks.
+
+    Args:
+        text (str): Input text.
+        drop_unvoiced_vowels (bool): whether to drop unvoiced vowels.
+
+    Returns:
+        List[str]: List of phoneme + prosody symbols.
+
+    Examples:
+        >>> from espnet2.text.phoneme_tokenizer import pyopenjtalk_g2p_prosody
+        >>> pyopenjtalk_g2p_prosody("こんにちは。")
+        ['^', 'k', 'o', '[', 'N', 'n', 'i', 'ch', 'i', 'w', 'a', '$']
+
+    .. _`Prosodic features control by symbols as input of sequence-to-sequence acoustic
+        modeling for neural TTS`: https://doi.org/10.1587/transinf.2020EDP7104
+
+    """
+    labels = pyopenjtalk.make_label(pyopenjtalk.run_frontend(text))
+    N = len(labels)
+
+    phones = []
+    for n in range(N):
+        lab_curr = labels[n]
+
+        # current phoneme
+        p3 = re.search(r"\-(.*?)\+", lab_curr).group(1)
+        # deal unvoiced vowels as normal vowels
+        if drop_unvoiced_vowels and p3 in "AEIOU":
+            p3 = p3.lower()
+
+        # deal with sil at the beginning and the end of text
+        if p3 == "sil":
+            assert n == 0 or n == N - 1
+            if n == 0:
+                phones.append("^")
+            elif n == N - 1:
+                # check question form or not
+                e3 = _numeric_feature_by_regex(r"!(\d+)_", lab_curr)
+                if e3 == 0:
+                    phones.append("$")
+                elif e3 == 1:
+                    phones.append("?")
+            continue
+        elif p3 == "pau":
+            phones.append("_")
+            continue
+        else:
+            phones.append(p3)
+
+        # accent type and position info (forward or backward)
+        a1 = _numeric_feature_by_regex(r"/A:([0-9\-]+)\+", lab_curr)
+        a2 = _numeric_feature_by_regex(r"\+(\d+)\+", lab_curr)
+        a3 = _numeric_feature_by_regex(r"\+(\d+)/", lab_curr)
+
+        # number of mora in accent phrase
+        f1 = _numeric_feature_by_regex(r"/F:(\d+)_", lab_curr)
+
+        a2_next = _numeric_feature_by_regex(r"\+(\d+)\+", labels[n + 1])
+        # accent phrase border
+        if a3 == 1 and a2_next == 1 and p3 in "aeiouAEIOUNcl":
+            phones.append("#")
+        # pitch falling
+        elif a1 == 0 and a2_next == a2 + 1 and a2 != f1:
+            phones.append("]")
+        # pitch rising
+        elif a2 == 1 and a2_next == 2:
+            phones.append("[")
+
+    return phones
+
+
+def _numeric_feature_by_regex(regex, s):
+    match = re.search(regex, s)
+    if match is None:
+        return -50
+    return int(match.group(1))
+
+
+def fix_phone_tone(phone_tone_list: list[tuple[str, int]]) -> list[tuple[str, int]]:
+    """
+    `phone_tone_list`のtone（アクセントの値）を0か1の範囲に修正する。
+    例: [(a, 0), (i, -1), (u, -1)] → [(a, 1), (i, 0), (u, 0)]
+    """
+    tone_values = set(tone for _, tone in phone_tone_list)
+    if len(tone_values) == 1:
+        assert tone_values == {0}, tone_values
+        return phone_tone_list
+    elif len(tone_values) == 2:
+        if tone_values == {0, 1}:
+            return phone_tone_list
+        elif tone_values == {-1, 0}:
+            return [
+                (letter, 0 if tone == -1 else 1) for letter, tone in phone_tone_list
+            ]
+        else:
+            raise ValueError(f"Unexpected tone values: {tone_values}")
+    else:
+        raise ValueError(f"Unexpected tone values: {tone_values}")
+
+
+def distribute_phone(n_phone: int, n_word: int) -> list[int]:
+    """
+    左から右に1ずつ振り分け、次にまた左から右に1ずつ増やし、というふうに、
+    音素の数`n_phone`を単語の数`n_word`に分配する。
+    """
+    phones_per_word = [0] * n_word
+    for _ in range(n_phone):
+        min_tasks = min(phones_per_word)
+        min_index = phones_per_word.index(min_tasks)
+        phones_per_word[min_index] += 1
+    return phones_per_word
+
+
+def handle_long(sep_phonemes: list[list[str]]) -> list[list[str]]:
+    for i in range(len(sep_phonemes)):
+        if sep_phonemes[i][0] == "ー":
+            sep_phonemes[i][0] = sep_phonemes[i - 1][-1]
+        if "ー" in sep_phonemes[i]:
+            for j in range(len(sep_phonemes[i])):
+                if sep_phonemes[i][j] == "ー":
+                    sep_phonemes[i][j] = sep_phonemes[i][j - 1][-1]
+    return sep_phonemes
+
+
+tokenizer = AutoTokenizer.from_pretrained("./bert/deberta-v2-large-japanese-char-wwm")
+
+
+def align_tones(
+    phones_with_punct: list[str], phone_tone_list: list[tuple[str, int]]
+) -> list[tuple[str, int]]:
+    """
+    例:
+    …私は、、そう思う。
+    phones_with_punct:
+    [".", ".", ".", "w", "a", "t", "a", "sh", "i", "w", "a", ",", ",", "s", "o", "o", "o", "m", "o", "u", "."]
+    phone_tone_list:
+    [("w", 0), ("a", 0), ("t", 1), ("a", 1), ("sh", 1), ("i", 1), ("w", 1), ("a", 1), ("_", 0), ("s", 0), ("o", 0), ("o", 1), ("o", 1), ("m", 1), ("o", 1), ("u", 0))]
+    Return:
+    [(".", 0), (".", 0), (".", 0), ("w", 0), ("a", 0), ("t", 1), ("a", 1), ("sh", 1), ("i", 1), ("w", 1), ("a", 1), (",", 0), (",", 0), ("s", 0), ("o", 0), ("o", 1), ("o", 1), ("m", 1), ("o", 1), ("u", 0), (".", 0)]
+    """
+    result: list[tuple[str, int]] = []
+    tone_index = 0
+    for phone in phones_with_punct:
+        if tone_index >= len(phone_tone_list):
+            # 余ったpunctuationがある場合 → (punctuation, 0)を追加
+            result.append((phone, 0))
+        elif phone == phone_tone_list[tone_index][0]:
+            # phone_tone_listの現在の音素と一致する場合 → toneをそこから取得、(phone, tone)を追加
+            result.append((phone, phone_tone_list[tone_index][1]))
+            # 探すindexを1つ進める
+            tone_index += 1
+        elif phone in punctuation:
+            # phoneがpunctuationの場合 → (phone, 0)を追加
+            result.append((phone, 0))
+        else:
+            logger.debug(f"phones: {phones_with_punct}")
+            logger.debug(f"phone_tone_list: {phone_tone_list}")
+            logger.debug(f"result: {result}")
+            logger.debug(f"tone_index: {tone_index}")
+            logger.debug(f"phone: {phone}")
+            raise ValueError(f"Unexpected phone: {phone}")
+    return result
+
+
+def kata2phoneme_list(text: str) -> list[str]:
+    """
+    原則カタカナの`text`を受け取り、それをそのままいじらずに音素記号のリストに変換。
+    注意点：
+    - punctuationが来た場合（punctuationが1文字の場合がありうる）、処理せず1文字のリストを返す
+    - 冒頭に続く「ー」はそのまま「ー」のままにする（`handle_long()`で処理される）
+    - 文中の「ー」は前の音素記号の最後の音素記号に変換される。
+    例：
+    `ーーソーナノカーー` → ["ー", "ー", "s", "o", "o", "n", "a", "n", "o", "k", "a", "a", "a"]
+    `?` → ["?"]
+    """
+    if text in punctuation:
+        return [text]
+    elif text == "--":
+        return ["-", "-"]
+    # `text`がカタカナ（`ー`含む）のみからなるかどうかをチェック
+    if re.fullmatch(r"[\u30A0-\u30FF]+", text) is None:
+        raise ValueError(f"Input must be katakana only: {text}")
+    sorted_keys = sorted(mora_kata_to_mora_phonemes.keys(), key=len, reverse=True)
+    pattern = "|".join(map(re.escape, sorted_keys))
+
+    def mora2phonemes(mora: str) -> str:
+        cosonant, vowel = mora_kata_to_mora_phonemes[mora]
+        if cosonant is None:
+            return f" {vowel}"
+        return f" {cosonant} {vowel}"
+
+    spaced_phonemes = re.sub(pattern, lambda m: mora2phonemes(m.group()), text)
+
+    # 長音記号「ー」の処理
+    long_pattern = r"(\w)(ー*)"
+    long_replacement = lambda m: m.group(1) + (" " + m.group(1)) * len(m.group(2))
+    spaced_phonemes = re.sub(long_pattern, long_replacement, spaced_phonemes)
+    return spaced_phonemes.strip().split(" ")
+
+
+if __name__ == "__main__":
+    tokenizer = AutoTokenizer.from_pretrained("./bert/deberta-v2-large-japanese")
+    text = "hello,こんにちは、世界ー！……"
+    from text.japanese_bert import get_bert_feature
+
+    text = text_normalize(text)
+    print(text)
+
+    phones, tones, word2ph = g2p(text)
+    bert = get_bert_feature(text, word2ph)
+
+    print(phones, tones, word2ph, bert.shape)

text/japanese_bert.py

@@ -0,0 +1,67 @@
+import sys
+
+import torch
+from transformers import AutoModelForMaskedLM, AutoTokenizer
+
+from config import config
+from text.japanese import text2sep_kata
+
+LOCAL_PATH = "./bert/deberta-v2-large-japanese-char-wwm"
+
+tokenizer = AutoTokenizer.from_pretrained(LOCAL_PATH)
+
+models = dict()
+
+
+def get_bert_feature(
+    text,
+    word2ph,
+    device=config.bert_gen_config.device,
+    assist_text=None,
+    assist_text_weight=0.7,
+):
+    text = "".join(text2sep_kata(text)[0])
+    if assist_text:
+        assist_text = "".join(text2sep_kata(assist_text)[0])
+    if (
+        sys.platform == "darwin"
+        and torch.backends.mps.is_available()
+        and device == "cpu"
+    ):
+        device = "mps"
+    if not device:
+        device = "cuda"
+    if device == "cuda" and not torch.cuda.is_available():
+        device = "cpu"
+    if device not in models.keys():
+        models[device] = AutoModelForMaskedLM.from_pretrained(LOCAL_PATH).to(device)
+    with torch.no_grad():
+        inputs = tokenizer(text, return_tensors="pt")
+        for i in inputs:
+            inputs[i] = inputs[i].to(device)
+        res = models[device](**inputs, output_hidden_states=True)
+        res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
+        if assist_text:
+            style_inputs = tokenizer(assist_text, return_tensors="pt")
+            for i in style_inputs:
+                style_inputs[i] = style_inputs[i].to(device)
+            style_res = models[device](**style_inputs, output_hidden_states=True)
+            style_res = torch.cat(style_res["hidden_states"][-3:-2], -1)[0].cpu()
+            style_res_mean = style_res.mean(0)
+
+    assert len(word2ph) == len(text) + 2, text
+    word2phone = word2ph
+    phone_level_feature = []
+    for i in range(len(word2phone)):
+        if assist_text:
+            repeat_feature = (
+                res[i].repeat(word2phone[i], 1) * (1 - assist_text_weight)
+                + style_res_mean.repeat(word2phone[i], 1) * assist_text_weight
+            )
+        else:
+            repeat_feature = res[i].repeat(word2phone[i], 1)
+        phone_level_feature.append(repeat_feature)
+
+    phone_level_feature = torch.cat(phone_level_feature, dim=0)
+
+    return phone_level_feature.T

text/japanese_mora_list.py

@@ -0,0 +1,232 @@
+"""
+VOICEVOXのソースコードからお借りして最低限に改造したコード。
+https://github.com/VOICEVOX/voicevox_engine/blob/master/voicevox_engine/tts_pipeline/mora_list.py
+"""
+
+"""
+以下のモーラ対応表はOpenJTalkのソースコードから取得し、
+カタカナ表記とモーラが一対一対応するように改造した。
+ライセンス表記：
+-----------------------------------------------------------------
+          The Japanese TTS System "Open JTalk"
+          developed by HTS Working Group
+          http://open-jtalk.sourceforge.net/
+-----------------------------------------------------------------
+
+ Copyright (c) 2008-2014  Nagoya Institute of Technology
+                          Department of Computer Science
+
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or
+without modification, are permitted provided that the following
+conditions are met:
+
+- Redistributions of source code must retain the above copyright
+  notice, this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above
+  copyright notice, this list of conditions and the following
+  disclaimer in the documentation and/or other materials provided
+  with the distribution.
+- Neither the name of the HTS working group nor the names of its
+  contributors may be used to endorse or promote products derived
+  from this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND
+CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES,
+INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS
+BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
+TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
+OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+"""
+from typing import Optional
+
+# (カタカナ, 子音, 母音)の順。子音がない場合はNoneを入れる。
+# 但し「ン」と「ッ」は母音のみという扱いで、「ン」は「N」、「ッ」は「q」とする。
+# （元々「ッ」は「cl」）
+# また「デェ = dy e」はpyopenjtalkの出力（de e）と合わないため削除
+_mora_list_minimum: list[tuple[str, Optional[str], str]] = [
+    ("ヴォ", "v", "o"),
+    ("ヴェ", "v", "e"),
+    ("ヴィ", "v", "i"),
+    ("ヴァ", "v", "a"),
+    ("ヴ", "v", "u"),
+    ("ン", None, "N"),
+    ("ワ", "w", "a"),
+    ("ロ", "r", "o"),
+    ("レ", "r", "e"),
+    ("ル", "r", "u"),
+    ("リョ", "ry", "o"),
+    ("リュ", "ry", "u"),
+    ("リャ", "ry", "a"),
+    ("リェ", "ry", "e"),
+    ("リ", "r", "i"),
+    ("ラ", "r", "a"),
+    ("ヨ", "y", "o"),
+    ("ユ", "y", "u"),
+    ("ヤ", "y", "a"),
+    ("モ", "m", "o"),
+    ("メ", "m", "e"),
+    ("ム", "m", "u"),
+    ("ミョ", "my", "o"),
+    ("ミュ", "my", "u"),
+    ("ミャ", "my", "a"),
+    ("ミェ", "my", "e"),
+    ("ミ", "m", "i"),
+    ("マ", "m", "a"),
+    ("ポ", "p", "o"),
+    ("ボ", "b", "o"),
+    ("ホ", "h", "o"),
+    ("ペ", "p", "e"),
+    ("ベ", "b", "e"),
+    ("ヘ", "h", "e"),
+    ("プ", "p", "u"),
+    ("ブ", "b", "u"),
+    ("フォ", "f", "o"),
+    ("フェ", "f", "e"),
+    ("フィ", "f", "i"),
+    ("ファ", "f", "a"),
+    ("フ", "f", "u"),
+    ("ピョ", "py", "o"),
+    ("ピュ", "py", "u"),
+    ("ピャ", "py", "a"),
+    ("ピェ", "py", "e"),
+    ("ピ", "p", "i"),
+    ("ビョ", "by", "o"),
+    ("ビュ", "by", "u"),
+    ("ビャ", "by", "a"),
+    ("ビェ", "by", "e"),
+    ("ビ", "b", "i"),
+    ("ヒョ", "hy", "o"),
+    ("ヒュ", "hy", "u"),
+    ("ヒャ", "hy", "a"),
+    ("ヒェ", "hy", "e"),
+    ("ヒ", "h", "i"),
+    ("パ", "p", "a"),
+    ("バ", "b", "a"),
+    ("ハ", "h", "a"),
+    ("ノ", "n", "o"),
+    ("ネ", "n", "e"),
+    ("ヌ", "n", "u"),
+    ("ニョ", "ny", "o"),
+    ("ニュ", "ny", "u"),
+    ("ニャ", "ny", "a"),
+    ("ニェ", "ny", "e"),
+    ("ニ", "n", "i"),
+    ("ナ", "n", "a"),
+    ("ドゥ", "d", "u"),
+    ("ド", "d", "o"),
+    ("トゥ", "t", "u"),
+    ("ト", "t", "o"),
+    ("デョ", "dy", "o"),
+    ("デュ", "dy", "u"),
+    ("デャ", "dy", "a"),
+    # ("デェ", "dy", "e"),
+    ("ディ", "d", "i"),
+    ("デ", "d", "e"),
+    ("テョ", "ty", "o"),
+    ("テュ", "ty", "u"),
+    ("テャ", "ty", "a"),
+    ("ティ", "t", "i"),
+    ("テ", "t", "e"),
+    ("ツォ", "ts", "o"),
+    ("ツェ", "ts", "e"),
+    ("ツィ", "ts", "i"),
+    ("ツァ", "ts", "a"),
+    ("ツ", "ts", "u"),
+    ("ッ", None, "q"),  # 「cl」から「q」に変更
+    ("チョ", "ch", "o"),
+    ("チュ", "ch", "u"),
+    ("チャ", "ch", "a"),
+    ("チェ", "ch", "e"),
+    ("チ", "ch", "i"),
+    ("ダ", "d", "a"),
+    ("タ", "t", "a"),
+    ("ゾ", "z", "o"),
+    ("ソ", "s", "o"),
+    ("ゼ", "z", "e"),
+    ("セ", "s", "e"),
+    ("ズィ", "z", "i"),
+    ("ズ", "z", "u"),
+    ("スィ", "s", "i"),
+    ("ス", "s", "u"),
+    ("ジョ", "j", "o"),
+    ("ジュ", "j", "u"),
+    ("ジャ", "j", "a"),
+    ("ジェ", "j", "e"),
+    ("ジ", "j", "i"),
+    ("ショ", "sh", "o"),
+    ("シュ", "sh", "u"),
+    ("シャ", "sh", "a"),
+    ("シェ", "sh", "e"),
+    ("シ", "sh", "i"),
+    ("ザ", "z", "a"),
+    ("サ", "s", "a"),
+    ("ゴ", "g", "o"),
+    ("コ", "k", "o"),
+    ("ゲ", "g", "e"),
+    ("ケ", "k", "e"),
+    ("グヮ", "gw", "a"),
+    ("グ", "g", "u"),
+    ("クヮ", "kw", "a"),
+    ("ク", "k", "u"),
+    ("ギョ", "gy", "o"),
+    ("ギュ", "gy", "u"),
+    ("ギャ", "gy", "a"),
+    ("ギェ", "gy", "e"),
+    ("ギ", "g", "i"),
+    ("キョ", "ky", "o"),
+    ("キュ", "ky", "u"),
+    ("キャ", "ky", "a"),
+    ("キェ", "ky", "e"),
+    ("キ", "k", "i"),
+    ("ガ", "g", "a"),
+    ("カ", "k", "a"),
+    ("オ", None, "o"),
+    ("エ", None, "e"),
+    ("ウォ", "w", "o"),
+    ("ウェ", "w", "e"),
+    ("ウィ", "w", "i"),
+    ("ウ", None, "u"),
+    ("イェ", "y", "e"),
+    ("イ", None, "i"),
+    ("ア", None, "a"),
+]
+_mora_list_additional: list[tuple[str, Optional[str], str]] = [
+    ("ヴョ", "by", "o"),
+    ("ヴュ", "by", "u"),
+    ("ヴャ", "by", "a"),
+    ("ヲ", None, "o"),
+    ("ヱ", None, "e"),
+    ("ヰ", None, "i"),
+    ("ヮ", "w", "a"),
+    ("ョ", "y", "o"),
+    ("ュ", "y", "u"),
+    ("ヅ", "z", "u"),
+    ("ヂ", "j", "i"),
+    ("ヶ", "k", "e"),
+    ("ャ", "y", "a"),
+    ("ォ", None, "o"),
+    ("ェ", None, "e"),
+    ("ゥ", None, "u"),
+    ("ィ", None, "i"),
+    ("ァ", None, "a"),
+]
+
+# 例: "vo" -> "ヴォ", "a" -> "ア"
+mora_phonemes_to_mora_kata: dict[str, str] = {
+    (consonant or "") + vowel: kana for [kana, consonant, vowel] in _mora_list_minimum
+}
+
+# 例: "ヴォ" -> ("v", "o"), "ア" -> (None, "a")
+mora_kata_to_mora_phonemes: dict[str, tuple[Optional[str], str]] = {
+    kana: (consonant, vowel)
+    for [kana, consonant, vowel] in _mora_list_minimum + _mora_list_additional
+}

text/opencpop-strict.txt

@@ -0,0 +1,429 @@
+a	AA a
+ai	AA ai
+an	AA an
+ang	AA ang
+ao	AA ao
+ba	b a
+bai	b ai
+ban	b an
+bang	b ang
+bao	b ao
+bei	b ei
+ben	b en
+beng	b eng
+bi	b i
+bian	b ian
+biao	b iao
+bie	b ie
+bin	b in
+bing	b ing
+bo	b o
+bu	b u
+ca	c a
+cai	c ai
+can	c an
+cang	c ang
+cao	c ao
+ce	c e
+cei	c ei
+cen	c en
+ceng	c eng
+cha	ch a
+chai	ch ai
+chan	ch an
+chang	ch ang
+chao	ch ao
+che	ch e
+chen	ch en
+cheng	ch eng
+chi	ch ir
+chong	ch ong
+chou	ch ou
+chu	ch u
+chua	ch ua
+chuai	ch uai
+chuan	ch uan
+chuang	ch uang
+chui	ch ui
+chun	ch un
+chuo	ch uo
+ci	c i0
+cong	c ong
+cou	c ou
+cu	c u
+cuan	c uan
+cui	c ui
+cun	c un
+cuo	c uo
+da	d a
+dai	d ai
+dan	d an
+dang	d ang
+dao	d ao
+de	d e
+dei	d ei
+den	d en
+deng	d eng
+di	d i
+dia	d ia
+dian	d ian
+diao	d iao
+die	d ie
+ding	d ing
+diu	d iu
+dong	d ong
+dou	d ou
+du	d u
+duan	d uan
+dui	d ui
+dun	d un
+duo	d uo
+e	EE e
+ei	EE ei
+en	EE en
+eng	EE eng
+er	EE er
+fa	f a
+fan	f an
+fang	f ang
+fei	f ei
+fen	f en
+feng	f eng
+fo	f o
+fou	f ou
+fu	f u
+ga	g a
+gai	g ai
+gan	g an
+gang	g ang
+gao	g ao
+ge	g e
+gei	g ei
+gen	g en
+geng	g eng
+gong	g ong
+gou	g ou
+gu	g u
+gua	g ua
+guai	g uai
+guan	g uan
+guang	g uang
+gui	g ui
+gun	g un
+guo	g uo
+ha	h a
+hai	h ai
+han	h an
+hang	h ang
+hao	h ao
+he	h e
+hei	h ei
+hen	h en
+heng	h eng
+hong	h ong
+hou	h ou
+hu	h u
+hua	h ua
+huai	h uai
+huan	h uan
+huang	h uang
+hui	h ui
+hun	h un
+huo	h uo
+ji	j i
+jia	j ia
+jian	j ian
+jiang	j iang
+jiao	j iao
+jie	j ie
+jin	j in
+jing	j ing
+jiong	j iong
+jiu	j iu
+ju	j v
+jv	j v
+juan	j van
+jvan	j van
+jue	j ve
+jve	j ve
+jun	j vn
+jvn	j vn
+ka	k a
+kai	k ai
+kan	k an
+kang	k ang
+kao	k ao
+ke	k e
+kei	k ei
+ken	k en
+keng	k eng
+kong	k ong
+kou	k ou
+ku	k u
+kua	k ua
+kuai	k uai
+kuan	k uan
+kuang	k uang
+kui	k ui
+kun	k un
+kuo	k uo
+la	l a
+lai	l ai
+lan	l an
+lang	l ang
+lao	l ao
+le	l e
+lei	l ei
+leng	l eng
+li	l i
+lia	l ia
+lian	l ian
+liang	l iang
+liao	l iao
+lie	l ie
+lin	l in
+ling	l ing
+liu	l iu
+lo	l o
+long	l ong
+lou	l ou
+lu	l u
+luan	l uan
+lun	l un
+luo	l uo
+lv	l v
+lve	l ve
+ma	m a
+mai	m ai
+man	m an
+mang	m ang
+mao	m ao
+me	m e
+mei	m ei
+men	m en
+meng	m eng
+mi	m i
+mian	m ian
+miao	m iao
+mie	m ie
+min	m in
+ming	m ing
+miu	m iu
+mo	m o
+mou	m ou
+mu	m u
+na	n a
+nai	n ai
+nan	n an
+nang	n ang
+nao	n ao
+ne	n e
+nei	n ei
+nen	n en
+neng	n eng
+ni	n i
+nian	n ian
+niang	n iang
+niao	n iao
+nie	n ie
+nin	n in
+ning	n ing
+niu	n iu
+nong	n ong
+nou	n ou
+nu	n u
+nuan	n uan
+nun	n un
+nuo	n uo
+nv	n v
+nve	n ve
+o	OO o
+ou	OO ou
+pa	p a
+pai	p ai
+pan	p an
+pang	p ang
+pao	p ao
+pei	p ei
+pen	p en
+peng	p eng
+pi	p i
+pian	p ian
+piao	p iao
+pie	p ie
+pin	p in
+ping	p ing
+po	p o
+pou	p ou
+pu	p u
+qi	q i
+qia	q ia
+qian	q ian
+qiang	q iang
+qiao	q iao
+qie	q ie
+qin	q in
+qing	q ing
+qiong	q iong
+qiu	q iu
+qu	q v
+qv	q v
+quan	q van
+qvan	q van
+que	q ve
+qve	q ve
+qun	q vn
+qvn	q vn
+ran	r an
+rang	r ang
+rao	r ao
+re	r e
+ren	r en
+reng	r eng
+ri	r ir
+rong	r ong
+rou	r ou
+ru	r u
+rua	r ua
+ruan	r uan
+rui	r ui
+run	r un
+ruo	r uo
+sa	s a
+sai	s ai
+san	s an
+sang	s ang
+sao	s ao
+se	s e
+sen	s en
+seng	s eng
+sha	sh a
+shai	sh ai
+shan	sh an
+shang	sh ang
+shao	sh ao
+she	sh e
+shei	sh ei
+shen	sh en
+sheng	sh eng
+shi	sh ir
+shou	sh ou
+shu	sh u
+shua	sh ua
+shuai	sh uai
+shuan	sh uan
+shuang	sh uang
+shui	sh ui
+shun	sh un
+shuo	sh uo
+si	s i0
+song	s ong
+sou	s ou
+su	s u
+suan	s uan
+sui	s ui
+sun	s un
+suo	s uo
+ta	t a
+tai	t ai
+tan	t an
+tang	t ang
+tao	t ao
+te	t e
+tei	t ei
+teng	t eng
+ti	t i
+tian	t ian
+tiao	t iao
+tie	t ie
+ting	t ing
+tong	t ong
+tou	t ou
+tu	t u
+tuan	t uan
+tui	t ui
+tun	t un
+tuo	t uo
+wa	w a
+wai	w ai
+wan	w an
+wang	w ang
+wei	w ei
+wen	w en
+weng	w eng
+wo	w o
+wu	w u
+xi	x i
+xia	x ia
+xian	x ian
+xiang	x iang
+xiao	x iao
+xie	x ie
+xin	x in
+xing	x ing
+xiong	x iong
+xiu	x iu
+xu	x v
+xv	x v
+xuan	x van
+xvan	x van
+xue	x ve
+xve	x ve
+xun	x vn
+xvn	x vn
+ya	y a
+yan	y En
+yang	y ang
+yao	y ao
+ye	y E
+yi	y i
+yin	y in
+ying	y ing
+yo	y o
+yong	y ong
+you	y ou
+yu	y v
+yv	y v
+yuan	y van
+yvan	y van
+yue	y ve
+yve	y ve
+yun	y vn
+yvn	y vn
+za	z a
+zai	z ai
+zan	z an
+zang	z ang
+zao	z ao
+ze	z e
+zei	z ei
+zen	z en
+zeng	z eng
+zha	zh a
+zhai	zh ai
+zhan	zh an
+zhang	zh ang
+zhao	zh ao
+zhe	zh e
+zhei	zh ei
+zhen	zh en
+zheng	zh eng
+zhi	zh ir
+zhong	zh ong
+zhou	zh ou
+zhu	zh u
+zhua	zh ua
+zhuai	zh uai
+zhuan	zh uan
+zhuang	zh uang
+zhui	zh ui
+zhun	zh un
+zhuo	zh uo
+zi	z i0
+zong	z ong
+zou	z ou
+zu	z u
+zuan	z uan
+zui	z ui
+zun	z un
+zuo	z uo

text/symbols.py

@@ -0,0 +1,187 @@
+punctuation = ["!", "?", "…", ",", ".", "'", "-"]
+pu_symbols = punctuation + ["SP", "UNK"]
+pad = "_"
+
+# chinese
+zh_symbols = [
+    "E",
+    "En",
+    "a",
+    "ai",
+    "an",
+    "ang",
+    "ao",
+    "b",
+    "c",
+    "ch",
+    "d",
+    "e",
+    "ei",
+    "en",
+    "eng",
+    "er",
+    "f",
+    "g",
+    "h",
+    "i",
+    "i0",
+    "ia",
+    "ian",
+    "iang",
+    "iao",
+    "ie",
+    "in",
+    "ing",
+    "iong",
+    "ir",
+    "iu",
+    "j",
+    "k",
+    "l",
+    "m",
+    "n",
+    "o",
+    "ong",
+    "ou",
+    "p",
+    "q",
+    "r",
+    "s",
+    "sh",
+    "t",
+    "u",
+    "ua",
+    "uai",
+    "uan",
+    "uang",
+    "ui",
+    "un",
+    "uo",
+    "v",
+    "van",
+    "ve",
+    "vn",
+    "w",
+    "x",
+    "y",
+    "z",
+    "zh",
+    "AA",
+    "EE",
+    "OO",
+]
+num_zh_tones = 6
+
+# japanese
+ja_symbols = [
+    "N",
+    "a",
+    "a:",
+    "b",
+    "by",
+    "ch",
+    "d",
+    "dy",
+    "e",
+    "e:",
+    "f",
+    "g",
+    "gy",
+    "h",
+    "hy",
+    "i",
+    "i:",
+    "j",
+    "k",
+    "ky",
+    "m",
+    "my",
+    "n",
+    "ny",
+    "o",
+    "o:",
+    "p",
+    "py",
+    "q",
+    "r",
+    "ry",
+    "s",
+    "sh",
+    "t",
+    "ts",
+    "ty",
+    "u",
+    "u:",
+    "w",
+    "y",
+    "z",
+    "zy",
+]
+num_ja_tones = 2
+
+# English
+en_symbols = [
+    "aa",
+    "ae",
+    "ah",
+    "ao",
+    "aw",
+    "ay",
+    "b",
+    "ch",
+    "d",
+    "dh",
+    "eh",
+    "er",
+    "ey",
+    "f",
+    "g",
+    "hh",
+    "ih",
+    "iy",
+    "jh",
+    "k",
+    "l",
+    "m",
+    "n",
+    "ng",
+    "ow",
+    "oy",
+    "p",
+    "r",
+    "s",
+    "sh",
+    "t",
+    "th",
+    "uh",
+    "uw",
+    "V",
+    "w",
+    "y",
+    "z",
+    "zh",
+]
+num_en_tones = 4
+
+# combine all symbols
+normal_symbols = sorted(set(zh_symbols + ja_symbols + en_symbols))
+symbols = [pad] + normal_symbols + pu_symbols
+sil_phonemes_ids = [symbols.index(i) for i in pu_symbols]
+
+# combine all tones
+num_tones = num_zh_tones + num_ja_tones + num_en_tones
+
+# language maps
+language_id_map = {"ZH": 0, "JP": 1, "EN": 2}
+num_languages = len(language_id_map.keys())
+
+language_tone_start_map = {
+    "ZH": 0,
+    "JP": num_zh_tones,
+    "EN": num_zh_tones + num_ja_tones,
+}
+
+if __name__ == "__main__":
+    a = set(zh_symbols)
+    b = set(en_symbols)
+    print(sorted(a & b))

text/tone_sandhi.py

@@ -0,0 +1,773 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import List
+from typing import Tuple
+
+import jieba
+from pypinyin import lazy_pinyin
+from pypinyin import Style
+
+
+class ToneSandhi:
+    def __init__(self):
+        self.must_neural_tone_words = {
+            "麻烦",
+            "麻利",
+            "鸳鸯",
+            "高粱",
+            "骨头",
+            "骆驼",
+            "马虎",
+            "首饰",
+            "馒头",
+            "馄饨",
+            "风筝",
+            "难为",
+            "队伍",
+            "阔气",
+            "闺女",
+            "门道",
+            "锄头",
+            "铺盖",
+            "铃铛",
+            "铁匠",
+            "钥匙",
+            "里脊",
+            "里头",
+            "部分",
+            "那么",
+            "道士",
+            "造化",
+            "迷糊",
+            "连累",
+            "这么",
+            "这个",
+            "运气",
+            "过去",
+            "软和",
+            "转悠",
+            "踏实",
+            "跳蚤",
+            "跟头",
+            "趔趄",
+            "财主",
+            "豆腐",
+            "讲究",
+            "记性",
+            "记号",
+            "认识",
+            "规矩",
+            "见识",
+            "裁缝",
+            "补丁",
+            "衣裳",
+            "衣服",
+            "衙门",
+            "街坊",
+            "行李",
+            "行当",
+            "蛤蟆",
+            "蘑菇",
+            "薄荷",
+            "葫芦",
+            "葡萄",
+            "萝卜",
+            "荸荠",
+            "苗条",
+            "苗头",
+            "苍蝇",
+            "芝麻",
+            "舒服",
+            "舒坦",
+            "舌头",
+            "自在",
+            "膏药",
+            "脾气",
+            "脑袋",
+            "脊梁",
+            "能耐",
+            "胳膊",
+            "胭脂",
+            "胡萝",
+            "胡琴",
+            "胡同",
+            "聪明",
+            "耽误",
+            "耽搁",
+            "耷拉",
+            "耳朵",
+            "老爷",
+            "老实",
+            "老婆",
+            "老头",
+            "老太",
+            "翻腾",
+            "罗嗦",
+            "罐头",
+            "编辑",
+            "结实",
+            "红火",
+            "累赘",
+            "糨糊",
+            "糊涂",
+            "精神",
+            "粮食",
+            "簸箕",
+            "篱笆",
+            "算计",
+            "算盘",
+            "答应",
+            "笤帚",
+            "笑语",
+            "笑话",
+            "窟窿",
+            "窝囊",
+            "窗户",
+            "稳当",
+            "稀罕",
+            "称呼",
+            "秧歌",
+            "秀气",
+            "秀才",
+            "福气",
+            "祖宗",
+            "砚台",
+            "码头",
+            "石榴",
+            "石头",
+            "石匠",
+            "知识",
+            "眼睛",
+            "眯缝",
+            "眨巴",
+            "眉毛",
+            "相声",
+            "盘算",
+            "白净",
+            "痢疾",
+            "痛快",
+            "疟疾",
+            "疙瘩",
+            "疏忽",
+            "畜生",
+            "生意",
+            "甘蔗",
+            "琵琶",
+            "琢磨",
+            "琉璃",
+            "玻璃",
+            "玫瑰",
+            "玄乎",
+            "狐狸",
+            "状元",
+            "特务",
+            "牲口",
+            "牙碜",
+            "牌楼",
+            "爽快",
+            "爱人",
+            "热闹",
+            "烧饼",
+            "烟筒",
+            "烂糊",
+            "点心",
+            "炊帚",
+            "灯笼",
+            "火候",
+            "漂亮",
+            "滑溜",
+            "溜达",
+            "温和",
+            "清楚",
+            "消息",
+            "浪头",
+            "活泼",
+            "比方",
+            "正经",
+            "欺负",
+            "模糊",
+            "槟榔",
+            "棺材",
+            "棒槌",
+            "棉花",
+            "核桃",
+            "栅栏",
+            "柴火",
+            "架势",
+            "枕头",
+            "枇杷",
+            "机灵",
+            "本事",
+            "木头",
+            "木匠",
+            "朋友",
+            "月饼",
+            "月亮",
+            "暖和",
+            "明白",
+            "时候",
+            "新鲜",
+            "故事",
+            "收拾",
+            "收成",
+            "提防",
+            "挖苦",
+            "挑剔",
+            "指甲",
+            "指头",
+            "拾掇",
+            "拳头",
+            "拨弄",
+            "招牌",
+            "招呼",
+            "抬举",
+            "护士",
+            "折腾",
+            "扫帚",
+            "打量",
+            "打算",
+            "打点",
+            "打扮",
+            "打听",
+            "打发",
+            "扎实",
+            "扁担",
+            "戒指",
+            "懒得",
+            "意识",
+            "意思",
+            "情形",
+            "悟性",
+            "怪物",
+            "思量",
+            "怎么",
+            "念头",
+            "念叨",
+            "快活",
+            "忙活",
+            "志气",
+            "心思",
+            "得罪",
+            "张罗",
+            "弟兄",
+            "开通",
+            "应酬",
+            "庄稼",
+            "干事",
+            "帮手",
+            "帐篷",
+            "希罕",
+            "师父",
+            "师傅",
+            "巴结",
+            "巴掌",
+            "差事",
+            "工夫",
+            "岁数",
+            "屁股",
+            "尾巴",
+            "少爷",
+            "小气",
+            "小伙",
+            "将就",
+            "对头",
+            "对付",
+            "寡妇",
+            "家伙",
+            "客气",
+            "实在",
+            "官司",
+            "学问",
+            "学生",
+            "字号",
+            "嫁妆",
+            "媳妇",
+            "媒人",
+            "婆家",
+            "娘家",
+            "委屈",
+            "姑娘",
+            "姐夫",
+            "妯娌",
+            "妥当",
+            "妖精",
+            "奴才",
+            "女婿",
+            "头发",
+            "太阳",
+            "大爷",
+            "大方",
+            "大意",
+            "大夫",
+            "多少",
+            "多么",
+            "外甥",
+            "壮实",
+            "地道",
+            "地方",
+            "在乎",
+            "困难",
+            "嘴巴",
+            "嘱咐",
+            "嘟囔",
+            "嘀咕",
+            "喜欢",
+            "喇嘛",
+            "喇叭",
+            "商量",
+            "唾沫",
+            "哑巴",
+            "哈欠",
+            "哆嗦",
+            "咳嗽",
+            "和尚",
+            "告诉",
+            "告示",
+            "含糊",
+            "吓唬",
+            "后头",
+            "名字",
+            "名堂",
+            "合同",
+            "吆喝",
+            "叫唤",
+            "口袋",
+            "厚道",
+            "厉害",
+            "千斤",
+            "包袱",
+            "包涵",
+            "匀称",
+            "勤快",
+            "动静",
+            "动弹",
+            "功夫",
+            "力气",
+            "前头",
+            "刺猬",
+            "刺激",
+            "别扭",
+            "利落",
+            "利索",
+            "利害",
+            "分析",
+            "出息",
+            "凑合",
+            "凉快",
+            "冷战",
+            "冤枉",
+            "冒失",
+            "养活",
+            "关系",
+            "先生",
+            "兄弟",
+            "便宜",
+            "使唤",
+            "佩服",
+            "作坊",
+            "体面",
+            "位置",
+            "似的",
+            "伙计",
+            "休息",
+            "什么",
+            "人家",
+            "亲戚",
+            "亲家",
+            "交情",
+            "云彩",
+            "事情",
+            "买卖",
+            "主意",
+            "丫头",
+            "丧气",
+            "两口",
+            "东西",
+            "东家",
+            "世故",
+            "不由",
+            "不在",
+            "下水",
+            "下巴",
+            "上头",
+            "上司",
+            "丈夫",
+            "丈人",
+            "一辈",
+            "那个",
+            "菩萨",
+            "父亲",
+            "母亲",
+            "咕噜",
+            "邋遢",
+            "费用",
+            "冤家",
+            "甜头",
+            "介绍",
+            "荒唐",
+            "大人",
+            "泥鳅",
+            "幸福",
+            "熟悉",
+            "计划",
+            "扑腾",
+            "蜡烛",
+            "姥爷",
+            "照顾",
+            "喉咙",
+            "吉他",
+            "弄堂",
+            "蚂蚱",
+            "凤凰",
+            "拖沓",
+            "寒碜",
+            "糟蹋",
+            "倒腾",
+            "报复",
+            "逻辑",
+            "盘缠",
+            "喽啰",
+            "牢骚",
+            "咖喱",
+            "扫把",
+            "惦记",
+        }
+        self.must_not_neural_tone_words = {
+            "男子",
+            "女子",
+            "分子",
+            "原子",
+            "量子",
+            "莲子",
+            "石子",
+            "瓜子",
+            "电子",
+            "人人",
+            "虎虎",
+        }
+        self.punc = "：，；。？！“”‘’':,;.?!"
+
+    # the meaning of jieba pos tag: https://blog.csdn.net/weixin_44174352/article/details/113731041
+    # e.g.
+    # word: "家里"
+    # pos: "s"
+    # finals: ['ia1', 'i3']
+    def _neural_sandhi(self, word: str, pos: str, finals: List[str]) -> List[str]:
+        # reduplication words for n. and v. e.g. 奶奶, 试试, 旺旺
+        for j, item in enumerate(word):
+            if (
+                j - 1 >= 0
+                and item == word[j - 1]
+                and pos[0] in {"n", "v", "a"}
+                and word not in self.must_not_neural_tone_words
+            ):
+                finals[j] = finals[j][:-1] + "5"
+        ge_idx = word.find("个")
+        if len(word) >= 1 and word[-1] in "吧呢啊呐噻嘛吖嗨呐哦哒额滴哩哟喽啰耶喔诶":
+            finals[-1] = finals[-1][:-1] + "5"
+        elif len(word) >= 1 and word[-1] in "的地得":
+            finals[-1] = finals[-1][:-1] + "5"
+        # e.g. 走了, 看着, 去过
+        # elif len(word) == 1 and word in "了着过" and pos in {"ul", "uz", "ug"}:
+        #     finals[-1] = finals[-1][:-1] + "5"
+        elif (
+            len(word) > 1
+            and word[-1] in "们子"
+            and pos in {"r", "n"}
+            and word not in self.must_not_neural_tone_words
+        ):
+            finals[-1] = finals[-1][:-1] + "5"
+        # e.g. 桌上, 地下, 家里
+        elif len(word) > 1 and word[-1] in "上下里" and pos in {"s", "l", "f"}:
+            finals[-1] = finals[-1][:-1] + "5"
+        # e.g. 上来, 下去
+        elif len(word) > 1 and word[-1] in "来去" and word[-2] in "上下进出回过起开":
+            finals[-1] = finals[-1][:-1] + "5"
+        # 个做量词
+        elif (
+            ge_idx >= 1
+            and (word[ge_idx - 1].isnumeric() or word[ge_idx - 1] in "几有两半多各整每做是")
+        ) or word == "个":
+            finals[ge_idx] = finals[ge_idx][:-1] + "5"
+        else:
+            if (
+                word in self.must_neural_tone_words
+                or word[-2:] in self.must_neural_tone_words
+            ):
+                finals[-1] = finals[-1][:-1] + "5"
+
+        word_list = self._split_word(word)
+        finals_list = [finals[: len(word_list[0])], finals[len(word_list[0]) :]]
+        for i, word in enumerate(word_list):
+            # conventional neural in Chinese
+            if (
+                word in self.must_neural_tone_words
+                or word[-2:] in self.must_neural_tone_words
+            ):
+                finals_list[i][-1] = finals_list[i][-1][:-1] + "5"
+        finals = sum(finals_list, [])
+        return finals
+
+    def _bu_sandhi(self, word: str, finals: List[str]) -> List[str]:
+        # e.g. 看不懂
+        if len(word) == 3 and word[1] == "不":
+            finals[1] = finals[1][:-1] + "5"
+        else:
+            for i, char in enumerate(word):
+                # "不" before tone4 should be bu2, e.g. 不怕
+                if char == "不" and i + 1 < len(word) and finals[i + 1][-1] == "4":
+                    finals[i] = finals[i][:-1] + "2"
+        return finals
+
+    def _yi_sandhi(self, word: str, finals: List[str]) -> List[str]:
+        # "一" in number sequences, e.g. 一零零, 二一零
+        if word.find("一") != -1 and all(
+            [item.isnumeric() for item in word if item != "一"]
+        ):
+            return finals
+        # "一" between reduplication words should be yi5, e.g. 看一看
+        elif len(word) == 3 and word[1] == "一" and word[0] == word[-1]:
+            finals[1] = finals[1][:-1] + "5"
+        # when "一" is ordinal word, it should be yi1
+        elif word.startswith("第一"):
+            finals[1] = finals[1][:-1] + "1"
+        else:
+            for i, char in enumerate(word):
+                if char == "一" and i + 1 < len(word):
+                    # "一" before tone4 should be yi2, e.g. 一段
+                    if finals[i + 1][-1] == "4":
+                        finals[i] = finals[i][:-1] + "2"
+                    # "一" before non-tone4 should be yi4, e.g. 一天
+                    else:
+                        # "一" 后面如果是标点，还读一声
+                        if word[i + 1] not in self.punc:
+                            finals[i] = finals[i][:-1] + "4"
+        return finals
+
+    def _split_word(self, word: str) -> List[str]:
+        word_list = jieba.cut_for_search(word)
+        word_list = sorted(word_list, key=lambda i: len(i), reverse=False)
+        first_subword = word_list[0]
+        first_begin_idx = word.find(first_subword)
+        if first_begin_idx == 0:
+            second_subword = word[len(first_subword) :]
+            new_word_list = [first_subword, second_subword]
+        else:
+            second_subword = word[: -len(first_subword)]
+            new_word_list = [second_subword, first_subword]
+        return new_word_list
+
+    def _three_sandhi(self, word: str, finals: List[str]) -> List[str]:
+        if len(word) == 2 and self._all_tone_three(finals):
+            finals[0] = finals[0][:-1] + "2"
+        elif len(word) == 3:
+            word_list = self._split_word(word)
+            if self._all_tone_three(finals):
+                #  disyllabic + monosyllabic, e.g. 蒙古/包
+                if len(word_list[0]) == 2:
+                    finals[0] = finals[0][:-1] + "2"
+                    finals[1] = finals[1][:-1] + "2"
+                #  monosyllabic + disyllabic, e.g. 纸/老虎
+                elif len(word_list[0]) == 1:
+                    finals[1] = finals[1][:-1] + "2"
+            else:
+                finals_list = [finals[: len(word_list[0])], finals[len(word_list[0]) :]]
+                if len(finals_list) == 2:
+                    for i, sub in enumerate(finals_list):
+                        # e.g. 所有/人
+                        if self._all_tone_three(sub) and len(sub) == 2:
+                            finals_list[i][0] = finals_list[i][0][:-1] + "2"
+                        # e.g. 好/喜欢
+                        elif (
+                            i == 1
+                            and not self._all_tone_three(sub)
+                            and finals_list[i][0][-1] == "3"
+                            and finals_list[0][-1][-1] == "3"
+                        ):
+                            finals_list[0][-1] = finals_list[0][-1][:-1] + "2"
+                        finals = sum(finals_list, [])
+        # split idiom into two words who's length is 2
+        elif len(word) == 4:
+            finals_list = [finals[:2], finals[2:]]
+            finals = []
+            for sub in finals_list:
+                if self._all_tone_three(sub):
+                    sub[0] = sub[0][:-1] + "2"
+                finals += sub
+
+        return finals
+
+    def _all_tone_three(self, finals: List[str]) -> bool:
+        return all(x[-1] == "3" for x in finals)
+
+    # merge "不" and the word behind it
+    # if don't merge, "不" sometimes appears alone according to jieba, which may occur sandhi error
+    def _merge_bu(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        last_word = ""
+        for word, pos in seg:
+            if last_word == "不":
+                word = last_word + word
+            if word != "不":
+                new_seg.append((word, pos))
+            last_word = word[:]
+        if last_word == "不":
+            new_seg.append((last_word, "d"))
+            last_word = ""
+        return new_seg
+
+    # function 1: merge "一" and reduplication words in it's left and right, e.g. "听","一","听" ->"听一听"
+    # function 2: merge single  "一" and the word behind it
+    # if don't merge, "一" sometimes appears alone according to jieba, which may occur sandhi error
+    # e.g.
+    # input seg: [('听', 'v'), ('一', 'm'), ('听', 'v')]
+    # output seg: [['听一听', 'v']]
+    def _merge_yi(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = [] * len(seg)
+        # function 1
+        i = 0
+        while i < len(seg):
+            word, pos = seg[i]
+            if (
+                i - 1 >= 0
+                and word == "一"
+                and i + 1 < len(seg)
+                and seg[i - 1][0] == seg[i + 1][0]
+                and seg[i - 1][1] == "v"
+            ):
+                new_seg[i - 1][0] = new_seg[i - 1][0] + "一" + new_seg[i - 1][0]
+                i += 2
+            else:
+                if (
+                    i - 2 >= 0
+                    and seg[i - 1][0] == "一"
+                    and seg[i - 2][0] == word
+                    and pos == "v"
+                ):
+                    continue
+                else:
+                    new_seg.append([word, pos])
+                i += 1
+        seg = [i for i in new_seg if len(i) > 0]
+        new_seg = []
+        # function 2
+        for i, (word, pos) in enumerate(seg):
+            if new_seg and new_seg[-1][0] == "一":
+                new_seg[-1][0] = new_seg[-1][0] + word
+            else:
+                new_seg.append([word, pos])
+        return new_seg
+
+    # the first and the second words are all_tone_three
+    def _merge_continuous_three_tones(
+        self, seg: List[Tuple[str, str]]
+    ) -> List[Tuple[str, str]]:
+        new_seg = []
+        sub_finals_list = [
+            lazy_pinyin(word, neutral_tone_with_five=True, style=Style.FINALS_TONE3)
+            for (word, pos) in seg
+        ]
+        assert len(sub_finals_list) == len(seg)
+        merge_last = [False] * len(seg)
+        for i, (word, pos) in enumerate(seg):
+            if (
+                i - 1 >= 0
+                and self._all_tone_three(sub_finals_list[i - 1])
+                and self._all_tone_three(sub_finals_list[i])
+                and not merge_last[i - 1]
+            ):
+                # if the last word is reduplication, not merge, because reduplication need to be _neural_sandhi
+                if (
+                    not self._is_reduplication(seg[i - 1][0])
+                    and len(seg[i - 1][0]) + len(seg[i][0]) <= 3
+                ):
+                    new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
+                    merge_last[i] = True
+                else:
+                    new_seg.append([word, pos])
+            else:
+                new_seg.append([word, pos])
+
+        return new_seg
+
+    def _is_reduplication(self, word: str) -> bool:
+        return len(word) == 2 and word[0] == word[1]
+
+    # the last char of first word and the first char of second word is tone_three
+    def _merge_continuous_three_tones_2(
+        self, seg: List[Tuple[str, str]]
+    ) -> List[Tuple[str, str]]:
+        new_seg = []
+        sub_finals_list = [
+            lazy_pinyin(word, neutral_tone_with_five=True, style=Style.FINALS_TONE3)
+            for (word, pos) in seg
+        ]
+        assert len(sub_finals_list) == len(seg)
+        merge_last = [False] * len(seg)
+        for i, (word, pos) in enumerate(seg):
+            if (
+                i - 1 >= 0
+                and sub_finals_list[i - 1][-1][-1] == "3"
+                and sub_finals_list[i][0][-1] == "3"
+                and not merge_last[i - 1]
+            ):
+                # if the last word is reduplication, not merge, because reduplication need to be _neural_sandhi
+                if (
+                    not self._is_reduplication(seg[i - 1][0])
+                    and len(seg[i - 1][0]) + len(seg[i][0]) <= 3
+                ):
+                    new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
+                    merge_last[i] = True
+                else:
+                    new_seg.append([word, pos])
+            else:
+                new_seg.append([word, pos])
+        return new_seg
+
+    def _merge_er(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        for i, (word, pos) in enumerate(seg):
+            if i - 1 >= 0 and word == "儿" and seg[i - 1][0] != "#":
+                new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
+            else:
+                new_seg.append([word, pos])
+        return new_seg
+
+    def _merge_reduplication(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        for i, (word, pos) in enumerate(seg):
+            if new_seg and word == new_seg[-1][0]:
+                new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
+            else:
+                new_seg.append([word, pos])
+        return new_seg
+
+    def pre_merge_for_modify(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        seg = self._merge_bu(seg)
+        try:
+            seg = self._merge_yi(seg)
+        except:
+            print("_merge_yi failed")
+        seg = self._merge_reduplication(seg)
+        seg = self._merge_continuous_three_tones(seg)
+        seg = self._merge_continuous_three_tones_2(seg)
+        seg = self._merge_er(seg)
+        return seg
+
+    def modified_tone(self, word: str, pos: str, finals: List[str]) -> List[str]:
+        finals = self._bu_sandhi(word, finals)
+        finals = self._yi_sandhi(word, finals)
+        finals = self._neural_sandhi(word, pos, finals)
+        finals = self._three_sandhi(word, finals)
+        return finals

tools/__init__.py

@@ -0,0 +1,3 @@
+"""
+工具包
+"""

tools/classify_language.py

@@ -0,0 +1,197 @@
+import regex as re
+
+try:
+    from config import config
+
+    LANGUAGE_IDENTIFICATION_LIBRARY = (
+        config.webui_config.language_identification_library
+    )
+except:
+    LANGUAGE_IDENTIFICATION_LIBRARY = "langid"
+
+module = LANGUAGE_IDENTIFICATION_LIBRARY.lower()
+
+langid_languages = [
+    "af",
+    "am",
+    "an",
+    "ar",
+    "as",
+    "az",
+    "be",
+    "bg",
+    "bn",
+    "br",
+    "bs",
+    "ca",
+    "cs",
+    "cy",
+    "da",
+    "de",
+    "dz",
+    "el",
+    "en",
+    "eo",
+    "es",
+    "et",
+    "eu",
+    "fa",
+    "fi",
+    "fo",
+    "fr",
+    "ga",
+    "gl",
+    "gu",
+    "he",
+    "hi",
+    "hr",
+    "ht",
+    "hu",
+    "hy",
+    "id",
+    "is",
+    "it",
+    "ja",
+    "jv",
+    "ka",
+    "kk",
+    "km",
+    "kn",
+    "ko",
+    "ku",
+    "ky",
+    "la",
+    "lb",
+    "lo",
+    "lt",
+    "lv",
+    "mg",
+    "mk",
+    "ml",
+    "mn",
+    "mr",
+    "ms",
+    "mt",
+    "nb",
+    "ne",
+    "nl",
+    "nn",
+    "no",
+    "oc",
+    "or",
+    "pa",
+    "pl",
+    "ps",
+    "pt",
+    "qu",
+    "ro",
+    "ru",
+    "rw",
+    "se",
+    "si",
+    "sk",
+    "sl",
+    "sq",
+    "sr",
+    "sv",
+    "sw",
+    "ta",
+    "te",
+    "th",
+    "tl",
+    "tr",
+    "ug",
+    "uk",
+    "ur",
+    "vi",
+    "vo",
+    "wa",
+    "xh",
+    "zh",
+    "zu",
+]
+
+
+def classify_language(text: str, target_languages: list = None) -> str:
+    if module == "fastlid" or module == "fasttext":
+        from fastlid import fastlid, supported_langs
+
+        classifier = fastlid
+        if target_languages != None:
+            target_languages = [
+                lang for lang in target_languages if lang in supported_langs
+            ]
+            fastlid.set_languages = target_languages
+    elif module == "langid":
+        import langid
+
+        classifier = langid.classify
+        if target_languages != None:
+            target_languages = [
+                lang for lang in target_languages if lang in langid_languages
+            ]
+            langid.set_languages(target_languages)
+    else:
+        raise ValueError(f"Wrong module {module}")
+
+    lang = classifier(text)[0]
+
+    return lang
+
+
+def classify_zh_ja(text: str) -> str:
+    for idx, char in enumerate(text):
+        unicode_val = ord(char)
+
+        # 检测日语字符
+        if 0x3040 <= unicode_val <= 0x309F or 0x30A0 <= unicode_val <= 0x30FF:
+            return "ja"
+
+        # 检测汉字字符
+        if 0x4E00 <= unicode_val <= 0x9FFF:
+            # 检查周围的字符
+            next_char = text[idx + 1] if idx + 1 < len(text) else None
+
+            if next_char and (
+                0x3040 <= ord(next_char) <= 0x309F or 0x30A0 <= ord(next_char) <= 0x30FF
+            ):
+                return "ja"
+
+    return "zh"
+
+
+def split_alpha_nonalpha(text, mode=1):
+    if mode == 1:
+        pattern = r"(?<=[\u4e00-\u9fff\u3040-\u30FF\d\s])(?=[\p{Latin}])|(?<=[\p{Latin}\s])(?=[\u4e00-\u9fff\u3040-\u30FF\d])"
+    elif mode == 2:
+        pattern = r"(?<=[\u4e00-\u9fff\u3040-\u30FF\s])(?=[\p{Latin}\d])|(?<=[\p{Latin}\d\s])(?=[\u4e00-\u9fff\u3040-\u30FF])"
+    else:
+        raise ValueError("Invalid mode. Supported modes are 1 and 2.")
+
+    return re.split(pattern, text)
+
+
+if __name__ == "__main__":
+    text = "这是一个测试文本"
+    print(classify_language(text))
+    print(classify_zh_ja(text))  # "zh"
+
+    text = "これはテストテキストです"
+    print(classify_language(text))
+    print(classify_zh_ja(text))  # "ja"
+
+    text = "vits和Bert-VITS2是tts模型。花费3days.花费3天。Take 3 days"
+
+    print(split_alpha_nonalpha(text, mode=1))
+    # output: ['vits', '和', 'Bert-VITS', '2是', 'tts', '模型。花费3', 'days.花费3天。Take 3 days']
+
+    print(split_alpha_nonalpha(text, mode=2))
+    # output: ['vits', '和', 'Bert-VITS2', '是', 'tts', '模型。花费', '3days.花费', '3', '天。Take 3 days']
+
+    text = "vits 和 Bert-VITS2 是 tts 模型。花费3days.花费3天。Take 3 days"
+    print(split_alpha_nonalpha(text, mode=1))
+    # output: ['vits ', '和 ', 'Bert-VITS', '2 ', '是 ', 'tts ', '模型。花费3', 'days.花费3天。Take ', '3 ', 'days']
+
+    text = "vits 和 Bert-VITS2 是 tts 模型。花费3days.花费3天。Take 3 days"
+    print(split_alpha_nonalpha(text, mode=2))
+    # output: ['vits ', '和 ', 'Bert-VITS2 ', '是 ', 'tts ', '模型。花费', '3days.花费', '3', '天。Take ', '3 ', 'days']

tools/sentence.py

@@ -0,0 +1,173 @@
+import logging
+
+import regex as re
+
+from tools.classify_language import classify_language, split_alpha_nonalpha
+
+
+def check_is_none(item) -> bool:
+    """none -> True, not none -> False"""
+    return (
+        item is None
+        or (isinstance(item, str) and str(item).isspace())
+        or str(item) == ""
+    )
+
+
+def markup_language(text: str, target_languages: list = None) -> str:
+    pattern = (
+        r"[\!\"\#\$\%\&\'\(\)\*\+\,\-\.\/\:\;\<\>\=\?\@\[\]\{\}\\\\\^\_\`"
+        r"\！？。＂＃＄％＆＇（）＊＋，－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､、〃》「」"
+        r"『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘\'\‛\“\”\„\‟…‧﹏.]+"
+    )
+    sentences = re.split(pattern, text)
+
+    pre_lang = ""
+    p = 0
+
+    if target_languages is not None:
+        sorted_target_languages = sorted(target_languages)
+        if sorted_target_languages in [["en", "zh"], ["en", "ja"], ["en", "ja", "zh"]]:
+            new_sentences = []
+            for sentence in sentences:
+                new_sentences.extend(split_alpha_nonalpha(sentence))
+            sentences = new_sentences
+
+    for sentence in sentences:
+        if check_is_none(sentence):
+            continue
+
+        lang = classify_language(sentence, target_languages)
+
+        if pre_lang == "":
+            text = text[:p] + text[p:].replace(
+                sentence, f"[{lang.upper()}]{sentence}", 1
+            )
+            p += len(f"[{lang.upper()}]")
+        elif pre_lang != lang:
+            text = text[:p] + text[p:].replace(
+                sentence, f"[{pre_lang.upper()}][{lang.upper()}]{sentence}", 1
+            )
+            p += len(f"[{pre_lang.upper()}][{lang.upper()}]")
+        pre_lang = lang
+        p += text[p:].index(sentence) + len(sentence)
+    text += f"[{pre_lang.upper()}]"
+
+    return text
+
+
+def split_by_language(text: str, target_languages: list = None) -> list:
+    pattern = (
+        r"[\!\"\#\$\%\&\'\(\)\*\+\,\-\.\/\:\;\<\>\=\?\@\[\]\{\}\\\\\^\_\`"
+        r"\！？\。＂＃＄％＆＇（）＊＋，－／：；＜＝＞＠［＼］＾＿｀｛｜｝～｟｠｢｣､、〃》「」"
+        r"『』【】〔〕〖〗〘〙〚〛〜〝〞〟〰〾〿–—‘\'\‛\“\”\„\‟…‧﹏.]+"
+    )
+    sentences = re.split(pattern, text)
+
+    pre_lang = ""
+    start = 0
+    end = 0
+    sentences_list = []
+
+    if target_languages is not None:
+        sorted_target_languages = sorted(target_languages)
+        if sorted_target_languages in [["en", "zh"], ["en", "ja"], ["en", "ja", "zh"]]:
+            new_sentences = []
+            for sentence in sentences:
+                new_sentences.extend(split_alpha_nonalpha(sentence))
+            sentences = new_sentences
+
+    for sentence in sentences:
+        if check_is_none(sentence):
+            continue
+
+        lang = classify_language(sentence, target_languages)
+
+        end += text[end:].index(sentence)
+        if pre_lang != "" and pre_lang != lang:
+            sentences_list.append((text[start:end], pre_lang))
+            start = end
+        end += len(sentence)
+        pre_lang = lang
+    sentences_list.append((text[start:], pre_lang))
+
+    return sentences_list
+
+
+def sentence_split(text: str, max: int) -> list:
+    pattern = r"[!(),—+\-.:;?？。，、；：]+"
+    sentences = re.split(pattern, text)
+    discarded_chars = re.findall(pattern, text)
+
+    sentences_list, count, p = [], 0, 0
+
+    # 按被分割的符号遍历
+    for i, discarded_chars in enumerate(discarded_chars):
+        count += len(sentences[i]) + len(discarded_chars)
+        if count >= max:
+            sentences_list.append(text[p : p + count].strip())
+            p += count
+            count = 0
+
+    # 加入最后剩余的文本
+    if p < len(text):
+        sentences_list.append(text[p:])
+
+    return sentences_list
+
+
+def sentence_split_and_markup(text, max=50, lang="auto", speaker_lang=None):
+    # 如果该speaker只支持一种语言
+    if speaker_lang is not None and len(speaker_lang) == 1:
+        if lang.upper() not in ["AUTO", "MIX"] and lang.lower() != speaker_lang[0]:
+            logging.debug(
+                f'lang "{lang}" is not in speaker_lang {speaker_lang},automatically set lang={speaker_lang[0]}'
+            )
+        lang = speaker_lang[0]
+
+    sentences_list = []
+    if lang.upper() != "MIX":
+        if max <= 0:
+            sentences_list.append(
+                markup_language(text, speaker_lang)
+                if lang.upper() == "AUTO"
+                else f"[{lang.upper()}]{text}[{lang.upper()}]"
+            )
+        else:
+            for i in sentence_split(text, max):
+                if check_is_none(i):
+                    continue
+                sentences_list.append(
+                    markup_language(i, speaker_lang)
+                    if lang.upper() == "AUTO"
+                    else f"[{lang.upper()}]{i}[{lang.upper()}]"
+                )
+    else:
+        sentences_list.append(text)
+
+    for i in sentences_list:
+        logging.debug(i)
+
+    return sentences_list
+
+
+if __name__ == "__main__":
+    text = "这几天心里颇不宁静。今晚在院子里坐着乘凉，忽然想起日日走过的荷塘，在这满月的光里，总该另有一番样子吧。月亮渐渐地升高了，墙外马路上孩子们的欢笑，已经听不见了；妻在屋里拍着闰儿，迷迷糊糊地哼着眠歌。我悄悄地披了大衫，带上门出去。"
+    print(markup_language(text, target_languages=None))
+    print(sentence_split(text, max=50))
+    print(sentence_split_and_markup(text, max=50, lang="auto", speaker_lang=None))
+
+    text = "你好，这是一段用来测试自动标注的文本。こんにちは,これは自動ラベリングのテスト用テキストです.Hello, this is a piece of text to test autotagging.你好！今天我们要介绍VITS项目，其重点是使用了GAN Duration predictor和transformer flow,并且接入了Bert模型来提升韵律。Bert embedding会在稍后介绍。"
+    print(split_by_language(text, ["zh", "ja", "en"]))
+
+    text = "vits和Bert-VITS2是tts模型。花费3days.花费3天。Take 3 days"
+
+    print(split_by_language(text, ["zh", "ja", "en"]))
+    # output: [('vits', 'en'), ('和', 'ja'), ('Bert-VITS', 'en'), ('2是', 'zh'), ('tts', 'en'), ('模型。花费3', 'zh'), ('days.', 'en'), ('花费3天。', 'zh'), ('Take 3 days', 'en')]
+
+    print(split_by_language(text, ["zh", "en"]))
+    # output: [('vits', 'en'), ('和', 'zh'), ('Bert-VITS', 'en'), ('2是', 'zh'), ('tts', 'en'), ('模型。花费3', 'zh'), ('days.', 'en'), ('花费3天。', 'zh'), ('Take 3 days', 'en')]
+
+    text = "vits 和 Bert-VITS2 是 tts 模型。花费 3 days. 花费 3天。Take 3 days"
+    print(split_by_language(text, ["zh", "en"]))
+    # output: [('vits ', 'en'), ('和 ', 'zh'), ('Bert-VITS2 ', 'en'), ('是 ', 'zh'), ('tts ', 'en'), ('模型。花费 ', 'zh'), ('3 days. ', 'en'), ('花费 3天。', 'zh'), ('Take 3 days', 'en')]

tools/translate.py

@@ -0,0 +1,61 @@
+"""
+翻译api
+"""
+from config import config
+
+import random
+import hashlib
+import requests
+
+
+def translate(Sentence: str, to_Language: str = "jp", from_Language: str = ""):
+    """
+    :param Sentence: 待翻译语句
+    :param from_Language: 待翻译语句语言
+    :param to_Language: 目标语言
+    :return: 翻译后语句 出错时返回None
+
+    常见语言代码：中文 zh 英语 en 日语 jp
+    """
+    appid = config.translate_config.app_key
+    key = config.translate_config.secret_key
+    if appid == "" or key == "":
+        return "请开发者在config.yml中配置app_key与secret_key"
+    url = "https://fanyi-api.baidu.com/api/trans/vip/translate"
+    texts = Sentence.splitlines()
+    outTexts = []
+    for t in texts:
+        if t != "":
+            # 签名计算 参考文档 https://api.fanyi.baidu.com/product/113
+            salt = str(random.randint(1, 100000))
+            signString = appid + t + salt + key
+            hs = hashlib.md5()
+            hs.update(signString.encode("utf-8"))
+            signString = hs.hexdigest()
+            if from_Language == "":
+                from_Language = "auto"
+            headers = {"Content-Type": "application/x-www-form-urlencoded"}
+            payload = {
+                "q": t,
+                "from": from_Language,
+                "to": to_Language,
+                "appid": appid,
+                "salt": salt,
+                "sign": signString,
+            }
+            # 发送请求
+            try:
+                response = requests.post(
+                    url=url, data=payload, headers=headers, timeout=3
+                )
+                response = response.json()
+                if "trans_result" in response.keys():
+                    result = response["trans_result"][0]
+                    if "dst" in result.keys():
+                        dst = result["dst"]
+                        outTexts.append(dst)
+            except Exception:
+                return Sentence
+        else:
+            outTexts.append(t)
+    return "\n".join(outTexts)

utils.py

@@ -13,7 +13,7 @@ from safetensors import safe_open
 from safetensors.torch import save_file
 from scipy.io.wavfile import read
 
-from tools.log import logger
+from common.log import logger
 
 MATPLOTLIB_FLAG = False
 
@@ -189,10 +189,11 @@ def summarize(
 
 
 def is_resuming(dir_path):
+    # JP-ExtraバージョンではDURがなくWDがあったり変わるため、Gのみで判断する
     g_list = glob.glob(os.path.join(dir_path, "G_*.pth"))
-    d_list = glob.glob(os.path.join(dir_path, "D_*.pth"))
-    dur_list = glob.glob(os.path.join(dir_path, "DUR_*.pth"))
-    return len(g_list) > 0 and len(d_list) > 0 and len(dur_list) > 0
+    # d_list = glob.glob(os.path.join(dir_path, "D_*.pth"))
+    # dur_list = glob.glob(os.path.join(dir_path, "DUR_*.pth"))
+    return len(g_list) > 0
 
 
 def latest_checkpoint_path(dir_path, regex="G_*.pth"):
@@ -348,7 +349,7 @@ def clean_checkpoints(path_to_models="logs/44k/", n_ckpts_to_keep=2, sort_by_tim
     ]
 
     def del_info(fn):
-        return logger.info(f".. Free up space by deleting ckpt {fn}")
+        return logger.info(f"Free up space by deleting ckpt {fn}")
 
     def del_routine(x):
         return [os.remove(x), del_info(x)]