app.py

import os
import torch
import librosa
import numpy as np
import gradio as gr
import pyopenjtalk
from util import preprocess_input, postprocess_phn, get_tokenizer, load_pitch_dict, get_pinyin

from espnet_model_zoo.downloader import ModelDownloader
from espnet2.bin.svs_inference import SingingGenerate


singer_embeddings = {
    "Model①(Chinese)-zh": {
        "singer1 (male)": 1,
        "singer2 (female)": 12,
        "singer3 (male)": 23,
        "singer4 (female)": 29,
        "singer5 (male)": 18,
        "singer6 (female)": 8,
        "singer7 (male)": 25,
        "singer8 (female)": 5,
        "singer9 (male)": 10,
        "singer10 (female)": 15,
    },
    "Model②(Multilingual)-zh": {
        "singer1 (male)": "resource/singer/singer_embedding_ace-1.npy",
        "singer2 (female)": "resource/singer/singer_embedding_ace-2.npy",
        "singer3 (male)": "resource/singer/singer_embedding_ace-3.npy",
        "singer4 (female)": "resource/singer/singer_embedding_ace-8.npy",
        "singer5 (male)": "resource/singer/singer_embedding_ace-7.npy",
        "singer6 (female)": "resource/singer/singer_embedding_itako.npy",
        "singer7 (male)": "resource/singer/singer_embedding_ofuton.npy",
        "singer8 (female)": "resource/singer/singer_embedding_kising_orange.npy",
        "singer9 (male)": "resource/singer/singer_embedding_m4singer_Tenor-1.npy",
        "singer10 (female)": "resource/singer/singer_embedding_m4singer_Alto-4.npy",
    },
    "Model②(Multilingual)-jp": {
        "singer1 (male)": "resource/singer/singer_embedding_ace-1.npy",
        "singer2 (female)": "resource/singer/singer_embedding_ace-2.npy",
        "singer3 (male)": "resource/singer/singer_embedding_ace-3.npy",
        "singer4 (female)": "resource/singer/singer_embedding_ace-8.npy",
        "singer5 (male)": "resource/singer/singer_embedding_ace-7.npy",
        "singer6 (female)": "resource/singer/singer_embedding_itako.npy",
        "singer7 (male)": "resource/singer/singer_embedding_ofuton.npy",
        "singer8 (female)": "resource/singer/singer_embedding_kising_orange.npy",
        "singer9 (male)": "resource/singer/singer_embedding_m4singer_Tenor-1.npy",
        "singer10 (female)": "resource/singer/singer_embedding_m4singer_Alto-4.npy",
    },
}

model_dict = {
    "Model①(Chinese)-zh": "espnet/aceopencpop_svs_visinger2_40singer_pretrain",
    "Model②(Multilingual)-zh": "espnet/mixdata_svs_visinger2_spkembed_lang_pretrained",
    "Model②(Multilingual)-jp": "espnet/mixdata_svs_visinger2_spkembed_lang_pretrained",
}

total_singers = list(singer_embeddings["Model②(Multilingual)-zh"].keys())

langs = {
    "zh": 2,
    "jp": 1,
}

predictor = torch.hub.load("South-Twilight/SingMOS:v0.2.0", "singing_ssl_mos", trust_repo=True)
exist_model = "Null"
svs = None

def gen_song(model_name, spk, texts, durs, pitchs):
    fs = 44100
    tempo = 120
    lang = model_name.split("-")[-1]
    PRETRAIN_MODEL = model_dict[model_name]
    if texts is None:
        return (fs, np.array([0.0])), "Error: No Text provided!"
    if durs is None:
        return (fs, np.array([0.0])), "Error: No Dur provided!"
    if pitchs is None:
        return (fs, np.array([0.0])), "Error: No Pitch provided!"

    # preprocess
    if lang == "zh":
        texts = preprocess_input(texts, "")
        text_list = get_pinyin(texts)
    elif lang == "jp":
        texts = preprocess_input(texts, " ")
        text_list = texts.strip().split()
    durs = preprocess_input(durs, " ")
    dur_list = durs.strip().split()
    pitchs = preprocess_input(pitchs, " ")
    pitch_list = pitchs.strip().split()

    if len(text_list) != len(dur_list):
        return (fs, np.array([0.0])), f"Error: len in text({len(text_list)}) mismatch with duration({len(dur_list)})!"
    if len(text_list) != len(pitch_list):
        return (fs, np.array([0.0])), f"Error: len in text({len(text_list)}) mismatch with pitch({len(pitch_list)})!"

    ## text to phoneme
    tokenizer = get_tokenizer(model_name, lang)
    sybs = []
    for text in text_list:
        if text == "AP" or text == "SP":
            rev = [text]
        elif text == "-" or text == "——":
            rev = [text]
        else:
            rev = tokenizer(text)
        if rev == False:
            return (fs, np.array([0.0])), f"Error: text `{text}` is invalid!"
        rev = postprocess_phn(rev, model_name, lang)
        phns = "_".join(rev)
        sybs.append(phns)

    pitch_dict = load_pitch_dict()

    labels = []
    notes = []
    st = 0
    pre_phn = ""
    for phns, dur, pitch in zip(sybs, dur_list, pitch_list):
        if phns == "-" or phns == "——":
            phns = pre_phn
        if pitch not in pitch_dict:
            return (fs, np.array([0.0])), f"Error: pitch `{pitch}` is invalid!"
        pitch = pitch_dict[pitch]
        phn_list = phns.split("_")
        lyric = "".join(phn_list)
        dur = float(dur)
        note = [st, st + dur, lyric, pitch, phns]
        st += dur
        notes.append(note)
        for phn in phn_list:
            labels.append(phn)
        pre_phn = labels[-1]

    phns_str = " ".join(labels)
    batch = {
        "score": (
            int(tempo),
            notes,
        ),
        "text": phns_str,
    }
    print(batch)
    # return (fs, np.array([0.0])), "success!"

    # Infer
    global exist_model
    global svs
    if exist_model == "Null" or exist_model != model_name:
        device = "cpu"
        # device = "cuda" if torch.cuda.is_available() else "cpu"
        d = ModelDownloader()
        pretrain_downloaded = d.download_and_unpack(PRETRAIN_MODEL)
        svs = SingingGenerate(
            train_config = pretrain_downloaded["train_config"],
            model_file = pretrain_downloaded["model_file"],
            device = device
        )
        exist_model = model_name
    if model_name == "Model①(Chinese)-zh":
        sid = np.array([singer_embeddings[model_name][spk]])
        output_dict = svs(batch, sids=sid)
    else:
        lid = np.array([langs[lang]])
        spk_embed = np.load(singer_embeddings[model_name][spk])
        output_dict = svs(batch, lids=lid, spembs=spk_embed)
    wav_info = output_dict["wav"].cpu().numpy()

    # mos prediction with sr=16k
    global predictor
    wav_mos = librosa.resample(wav_info, orig_sr=fs, target_sr=16000)
    wav_mos = torch.from_numpy(wav_mos).unsqueeze(0)
    len_mos = torch.tensor([wav_mos.shape[1]])
    score = predictor(wav_mos, len_mos)
    return (fs, wav_info), "success!", round(score.item(), 2)


# SP: silence, AP: aspirate.
examples = [
    ["Model①(Chinese)-zh", "singer1 (male)", "雨 淋 湿 了 SP 天 空 AP\n毁 的 SP 很 讲 究 AP", "0.23 0.16 0.36 0.16 0.07 0.28 0.5 0.21\n0.3 0.12 0.12 0.25 0.5 0.48 0.34", "60 62 62 62 0 62 58 0\n58 58 0 58 58 63 0"],
    ["Model①(Chinese)-zh", "singer3 (male)", "雨 淋 湿 了 SP 天 空 AP\n毁 的 SP 很 讲 究 AP", "0.23 0.16 0.36 0.16 0.07 0.28 0.5 0.21\n0.3 0.12 0.12 0.25 0.5 0.48 0.34", "C4 D4 D4 D4 rest D4 A#3 rest\nA#3 A#3 rest A#3 A#3 D#4 rest"], # midi note
    ["Model①(Chinese)-zh", "singer3 (male)", "雨 淋 湿 了 SP 天 空 AP\n毁 的 SP 很 讲 究 AP", "0.23 0.16 0.36 0.16 0.07 0.28 0.5 0.21\n0.3 0.12 0.12 0.25 0.5 0.48 0.34", "C#4 D#4 D#4 D#4 rest D#4 B3 rest\nB3 B3 rest B3 B3 E4 rest"], # up 1 key
    ["Model①(Chinese)-zh", "singer3 (male)", "雨 淋 湿 了 SP 大 地 AP\n毁 的 SP 很 讲 究 AP", "0.23 0.16 0.36 0.16 0.07 0.28 0.5 0.21\n0.3 0.12 0.12 0.25 0.5 0.48 0.34", "C4 D4 D4 D4 rest D4 A#3 rest\nA#3 A#3 rest A#3 A#3 D#4 rest"], # lyrics
    ["Model②(Multilingual)-zh", "singer3 (male)", "你 说 你 不 SP 懂\n 为 何 在 这 时 牵 手 AP", "0.11 0.33 0.29 0.13 0.15 0.48\n0.24 0.18 0.34 0.15 0.27 0.28 0.63 0.44", "63 63 63 63 0 63\n62 62 62 63 65 63 62 0"],
    ["Model②(Multilingual)-zh", "singer3 (male)", "你 说 你 不 SP 懂\n 为 何 在 这 时 牵 手 AP", "0.23 0.66 0.58 0.27 0.3 0.97\n0.48 0.36 0.69 0.3 0.53 0.56 1.27 0.89", "63 63 63 63 0 63\n62 62 62 63 65 63 62 0"], # double duration
    ["Model①(Chinese)-zh", "singer3 (male)", "雨 淋 湿 了 SP 天 空 AP\n毁 的 SP 很 讲 究 AP\n你 说 你 不 SP 懂\n 为 何 在 这 时 牵 手 AP", "0.23 0.16 0.36 0.16 0.07 0.28 0.5 0.21\n0.3 0.12 0.12 0.25 0.5 0.48 0.34\n0.11 0.33 0.29 0.13 0.15 0.48\n0.24 0.18 0.34 0.15 0.27 0.28 0.63 0.44", "60 62 62 62 0 62 58 0\n58 58 0 58 58 63 0\n63 63 63 63 0 63\n62 62 62 63 65 63 62 0"], # long
    ["Model①(Chinese)-zh", "singer3 (male)", "修 炼 爱 情 的 心 酸 SP AP", "0.42 0.21 0.19 0.28 0.22 0.33 1.53 0.1 0.29", "68 70 68 66 63 68 68 0 0"],
    ["Model①(Chinese)-zh", "singer3 (male)", "学 会 放 好 以 前 的 渴 望 SP AP", "0.3 0.22 0.29 0.27 0.25 0.44 0.54 0.29 1.03 0.08 0.39", "68 70 68 66 61 68 68 65 66 0 0"],
    ["Model①(Chinese)-zh", "singer3 (male)", "SP 我 不 - 是 一 定 要 你 回 - 来 SP", "0.37 0.45 0.47 0.17 0.52 0.28 0.46 0.31 0.44 0.45 0.2 2.54 0.19", "0 51 60 61 59 59 57 57 59 60 61 59 0"], # slur
    ["Model①(Chinese)-zh", "singer4 (female)", "AP 我 多 想 再 见 你\n哪 怕 匆 - 匆 一 AP 眼 就 别 离 AP", "0.13 0.24 0.68 0.78 0.86 0.4 0.94 0.54 0.3 0.56 0.16 0.86 0.26 0.22 0.28 0.78 0.68 1.5 0.32", "0 57 66 63 63 63 63 60 61 61 63 66 66 0 61 61 59 58 0"],
    ["Model②(Multilingual)-jp", "singer8 (female)", "い じ ん さ ん に つ れ ら れ て", "0.6 0.3 0.3 0.3 0.3 0.6 0.6 0.3 0.3 0.6 0.23", "60 60 60 56 56 56 55 55 55 53 56"],
    ["Model②(Multilingual)-jp", "singer8 (female)", "い じ ん さ ん に つ れ ら れ て", "0.6 0.3 0.3 0.3 0.3 0.6 0.6 0.3 0.3 0.6 0.23", "62 62 62 58 58 58 57 57 57 55 58"], # pitch
    ["Model②(Multilingual)-jp", "singer8 (female)", "い じ ん さ ん に つ れ ら れ て", "1.2 0.6 0.6 0.6 0.6 1.2 1.2 0.6 0.6 1.2 0.45", "60 60 60 56 56 56 55 55 55 53 56"], # double dur
    ["Model②(Multilingual)-jp", "singer8 (female)", "い じ ん さ ん に つ れ ら れ て", "0.3 0.15 0.15 0.15 0.15 0.3 0.3 0.15 0.15 0.3 0.11", "60 60 60 56 56 56 55 55 55 53 56"], # half dur
    ["Model②(Multilingual)-jp", "singer8 (female)", "きっ と と べ ば そ ら ま で と ど く AP", "0.39 2.76 0.2 0.2 0.39 0.39 0.2 0.2 0.39 0.2 0.2 0.59 1.08", "64 71 68 69 71 71 69 68 66 68 69 68 0"],
    ["Model②(Multilingual)-jp", "singer8 (female)", "じゃ の め で お む か え う れ し い な", "0.43 0.14 0.43 0.14 0.43 0.14 0.43 0.14 0.43 0.14 0.43 0.14 0.65", "60 60 60 62 64 67 69 69 64 64 64 62 60"],
    ["Model②(Multilingual)-jp", "singer10 (female)", "お と め わ ら い か ふぁ い や ら い か ん な い す ぶ ろ うぃ ん ぶ ろ うぃ ん い ん ざ うぃ ん", "0.15 0.15 0.15 0.15 0.3 0.15 0.3 0.15 0.15 0.3 0.07 0.07 0.15 0.15 0.15 0.15 0.15 0.15 0.45 0.07 0.07 0.07 0.38 0.07 0.07 0.15 0.15 0.3 0.15 0.15", "67 67 67 67 67 67 69 67 67 69 67 67 64 64 64 64 64 64 62 64 64 62 62 64 64 62 62 59 59 59"],
]

with gr.Blocks() as demo:
    gr.Markdown(
        """
<h1 align="center"> Demo of Singing Voice Synthesis in Muskits-ESPnet </h1>

<div style="font-size: 20px;">
This is the demo page of our toolkit <a href="https://arxiv.org/abs/2409.07226"><b>Muskits-ESPnet: A Comprehensive Toolkit for Singing Voice Synthesis in New Paradigm</b></a>.

Singing Voice Synthesis (SVS) takes a music score as input and generates singing vocal with the voice of a specific singer.

Music score usually includes lyrics, as well as duration and pitch of each word in lyrics.

<h2>How to use:</h2>
    <ol>
        <li><b>Choose Model-Language</b>:
            <ul>
                <li>Choose "zh" for Chinese lyrics input or "jp" for Japanese lyrics input.</li>
                <li>For example, "Model②(Mulitlingual)-zh" means model "Model②(Multilingual)" with lyrics input in Chinese.</li>
            </ul>
        </li>
        <li><b>[Optional] Choose Singer</b>: Choose a singer from the drop-down menu.</li>
        <li><b>Input lyrics</b>:
            <ul>
                <li>Input Chinese characters for "zh" and hiragana for "jp".</li>
                <li>You may include special symbols: 'AP' for breath, 'SP' for silence, and '-' for slur (Chinese lyrics only).</li>
                <li>Separate each lyric by either a space (' ') or a newline ('\\n') (no quotation marks needed).</li>
            </ul>
        </li>
        <li><b>Input durations</b>:
            <ul>
                <li>Input durations as float numbers.</li>
                <li>The durations sequence should <b>match the lyric sequence in length</b>, with each duration aligned to a lyric.</li>
                <li>Separate each duration by a space (' ') or a newline ('\\n') (no quotation marks needed).</li>
            </ul>
        </li>
        <li><b>Input pitches</b>:
            <ul>
                <li>Input MIDI note names or MIDI note numbers (e.g., MIDI note name "69" represents the MIDI note number "A4", and others follow accordingly).</li>
                <li>The pitch sequence should <b>match the lyric sequence in length</b>, with each pitch corresponding to a lyric.</li>
                <li>Separate each duration by a space (' ') or a newline ('\\n') (no quotation marks needed).</li>
            </ul>
        </li>
        <li><b>Hit "Generate" and listen</b>:
            <ul>
                <li>"Running Status" shows the status of singing generatation. If any error exists, it will show the error information.</li>
                <li>"Pseudo MOS" represents predicted mean opinion score for the generated song.</li>
            </ul>
        </li>
    </ol>
</div>

<h2>Notice:</h2>
    <ul>
        <li> Plenty of exmpales are provided. </li>
        <li> Extreme values may result in suboptimal generation quality! </li>
    </ul>
"""       
    )
    # Row-1
    with gr.Row():
        with gr.Column(variant="panel"):
            model_name = gr.Radio(
                label="Model-Language",
                choices=[
                    "Model①(Chinese)-zh",
                    "Model②(Multilingual)-zh",
                    "Model②(Multilingual)-jp",
                ],
            )

        with gr.Column(variant="panel"):
            singer = gr.Dropdown(
                label="Singer",
                choices=total_singers,
            )
        
        # def set_model(model_name_str: str):
        #     """
        #     gets value from `model_name`. either
        #     uses cached list of speakers for the given model name
        #     or loads the addon and checks what are the speakers.
        #     """
        #     speakers = list(singer_embeddings[model_name_str].keys())
        #     value = speakers[0]
        #     return gr.update(
        #         choices=speakers, value=value, visible=True, interactive=True
        #     )

        # model_name.change(set_model, inputs=model_name, outputs=singer)

    # Row-2
    with gr.Row():
        with gr.Column(variant="panel"):
            lyrics = gr.Textbox(label="Lyrics")
            duration = gr.Textbox(label="Duration")
            pitch = gr.Textbox(label="Pitch")
            generate = gr.Button("Generate")
        with gr.Column(variant="panel"):
            gened_song = gr.Audio(label="Generated Song", type="numpy")
            run_status = gr.Textbox(label="Running Status")
            pred_mos = gr.Textbox(label=" Pseudo MOS")

    gr.Examples(
        examples=examples,
        inputs=[model_name, singer, lyrics, duration, pitch],
        outputs=[singer],
        label="Examples",
        examples_per_page=20,
    )
    
    gr.Markdown("""
<div style='margin:20px auto;'>

<p>References: <a href="https://arxiv.org/abs/2409.07226">Muskits-ESPnet paper</a> |
<a href="https://github.com/espnet/espnet">espnet</a> |
<a href="https://huggingface.co/espnet/aceopencpop_svs_visinger2_40singer_pretrain">Model①(Chinese)</a> |
<a href="https://huggingface.co/espnet/mixdata_svs_visinger2_spkembed_lang_pretrained">Model②(Multilingual)</a> |
<a href="https://github.com/South-Twilight/SingMOS">SingMOS</a></p>

</div>
"""
    )

    generate.click(
        fn=gen_song,
        inputs=[model_name, singer, lyrics, duration, pitch],
        outputs=[gened_song, run_status, pred_mos],
    )
    
demo.launch()