File size: 12,766 Bytes
f643c3e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
import os

gpt_path = os.environ.get(
    "gpt_path", "pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt"
)
sovits_path = os.environ.get("sovits_path", "pretrained_models/s2G488k.pth")
cnhubert_base_path = os.environ.get(
    "cnhubert_base_path", "pretrained_models/chinese-hubert-base"
)
bert_path = os.environ.get(
    "bert_path", "pretrained_models/chinese-roberta-wwm-ext-large"
)
infer_ttswebui = os.environ.get("infer_ttswebui", 9872)
infer_ttswebui = int(infer_ttswebui)
if "_CUDA_VISIBLE_DEVICES" in os.environ:
    os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["_CUDA_VISIBLE_DEVICES"]
is_half = eval(os.environ.get("is_half", "True"))
import gradio as gr
from transformers import AutoModelForMaskedLM, AutoTokenizer
import numpy as np
import librosa,torch
from feature_extractor import cnhubert
cnhubert.cnhubert_base_path=cnhubert_base_path

from module.models import SynthesizerTrn
from AR.models.t2s_lightning_module import Text2SemanticLightningModule
from text import cleaned_text_to_sequence
from text.cleaner import clean_text
from time import time as ttime
from module.mel_processing import spectrogram_torch
from my_utils import load_audio

device = "cuda"
tokenizer = AutoTokenizer.from_pretrained(bert_path)
bert_model = AutoModelForMaskedLM.from_pretrained(bert_path)
if is_half == True:
    bert_model = bert_model.half().to(device)
else:
    bert_model = bert_model.to(device)


# bert_model=bert_model.to(device)
def get_bert_feature(text, word2ph):
    with torch.no_grad():
        inputs = tokenizer(text, return_tensors="pt")
        for i in inputs:
            inputs[i] = inputs[i].to(device)  #####输入是long不用管精度问题,精度随bert_model
        res = bert_model(**inputs, output_hidden_states=True)
        res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()[1:-1]
    assert len(word2ph) == len(text)
    phone_level_feature = []
    for i in range(len(word2ph)):
        repeat_feature = res[i].repeat(word2ph[i], 1)
        phone_level_feature.append(repeat_feature)
    phone_level_feature = torch.cat(phone_level_feature, dim=0)
    # if(is_half==True):phone_level_feature=phone_level_feature.half()
    return phone_level_feature.T


n_semantic = 1024

dict_s2=torch.load(sovits_path,map_location="cpu")
hps=dict_s2["config"]

class DictToAttrRecursive(dict):
    def __init__(self, input_dict):
        super().__init__(input_dict)
        for key, value in input_dict.items():
            if isinstance(value, dict):
                value = DictToAttrRecursive(value)
            self[key] = value
            setattr(self, key, value)

    def __getattr__(self, item):
        try:
            return self[item]
        except KeyError:
            raise AttributeError(f"Attribute {item} not found")

    def __setattr__(self, key, value):
        if isinstance(value, dict):
            value = DictToAttrRecursive(value)
        super(DictToAttrRecursive, self).__setitem__(key, value)
        super().__setattr__(key, value)

    def __delattr__(self, item):
        try:
            del self[item]
        except KeyError:
            raise AttributeError(f"Attribute {item} not found")


hps = DictToAttrRecursive(hps)

hps.model.semantic_frame_rate = "25hz"
dict_s1 = torch.load(gpt_path, map_location="cpu")
config = dict_s1["config"]
ssl_model = cnhubert.get_model()
if is_half == True:
    ssl_model = ssl_model.half().to(device)
else:
    ssl_model = ssl_model.to(device)

vq_model = SynthesizerTrn(
    hps.data.filter_length // 2 + 1,
    hps.train.segment_size // hps.data.hop_length,
    n_speakers=hps.data.n_speakers,
    **hps.model
)
if is_half == True:
    vq_model = vq_model.half().to(device)
else:
    vq_model = vq_model.to(device)
vq_model.eval()
print(vq_model.load_state_dict(dict_s2["weight"], strict=False))
hz = 50
max_sec = config["data"]["max_sec"]
# t2s_model = Text2SemanticLightningModule.load_from_checkpoint(checkpoint_path=gpt_path, config=config, map_location="cpu")#########todo
t2s_model = Text2SemanticLightningModule(config, "ojbk", is_train=False)
t2s_model.load_state_dict(dict_s1["weight"])
if is_half == True:
    t2s_model = t2s_model.half()
t2s_model = t2s_model.to(device)
t2s_model.eval()
total = sum([param.nelement() for param in t2s_model.parameters()])
print("Number of parameter: %.2fM" % (total / 1e6))


def get_spepc(hps, filename):
    audio = load_audio(filename, int(hps.data.sampling_rate))
    audio = torch.FloatTensor(audio)
    audio_norm = audio
    audio_norm = audio_norm.unsqueeze(0)
    spec = spectrogram_torch(
        audio_norm,
        hps.data.filter_length,
        hps.data.sampling_rate,
        hps.data.hop_length,
        hps.data.win_length,
        center=False,
    )
    return spec


dict_language = {"中文": "zh", "英文": "en", "日文": "ja"}


def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language):
    t0 = ttime()
    prompt_text = prompt_text.strip("\n")
    prompt_language, text = prompt_language, text.strip("\n")
    with torch.no_grad():
        wav16k, sr = librosa.load(ref_wav_path, sr=16000)  # 派蒙
        wav16k = torch.from_numpy(wav16k)
        if is_half == True:
            wav16k = wav16k.half().to(device)
        else:
            wav16k = wav16k.to(device)
        ssl_content = ssl_model.model(wav16k.unsqueeze(0))[
            "last_hidden_state"
        ].transpose(
            1, 2
        )  # .float()
        codes = vq_model.extract_latent(ssl_content)
        prompt_semantic = codes[0, 0]
    t1 = ttime()
    prompt_language = dict_language[prompt_language]
    text_language = dict_language[text_language]
    phones1, word2ph1, norm_text1 = clean_text(prompt_text, prompt_language)
    phones1 = cleaned_text_to_sequence(phones1)
    texts = text.split("\n")
    audio_opt = []
    zero_wav = np.zeros(
        int(hps.data.sampling_rate * 0.3),
        dtype=np.float16 if is_half == True else np.float32,
    )
    for text in texts:
        # 解决输入目标文本的空行导致报错的问题
        if (len(text.strip()) == 0):
            continue
        phones2, word2ph2, norm_text2 = clean_text(text, text_language)
        phones2 = cleaned_text_to_sequence(phones2)
        if prompt_language == "zh":
            bert1 = get_bert_feature(norm_text1, word2ph1).to(device)
        else:
            bert1 = torch.zeros(
                (1024, len(phones1)),
                dtype=torch.float16 if is_half == True else torch.float32,
            ).to(device)
        if text_language == "zh":
            bert2 = get_bert_feature(norm_text2, word2ph2).to(device)
        else:
            bert2 = torch.zeros((1024, len(phones2))).to(bert1)
        bert = torch.cat([bert1, bert2], 1)

        all_phoneme_ids = torch.LongTensor(phones1 + phones2).to(device).unsqueeze(0)
        bert = bert.to(device).unsqueeze(0)
        all_phoneme_len = torch.tensor([all_phoneme_ids.shape[-1]]).to(device)
        prompt = prompt_semantic.unsqueeze(0).to(device)
        t2 = ttime()
        with torch.no_grad():
            # pred_semantic = t2s_model.model.infer(
            pred_semantic, idx = t2s_model.model.infer_panel(
                all_phoneme_ids,
                all_phoneme_len,
                prompt,
                bert,
                # prompt_phone_len=ph_offset,
                top_k=config["inference"]["top_k"],
                early_stop_num=hz * max_sec,
            )
        t3 = ttime()
        # print(pred_semantic.shape,idx)
        pred_semantic = pred_semantic[:, -idx:].unsqueeze(
            0
        )  # .unsqueeze(0)#mq要多unsqueeze一次
        refer = get_spepc(hps, ref_wav_path)  # .to(device)
        if is_half == True:
            refer = refer.half().to(device)
        else:
            refer = refer.to(device)
        # audio = vq_model.decode(pred_semantic, all_phoneme_ids, refer).detach().cpu().numpy()[0, 0]
        audio = (
            vq_model.decode(
                pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refer
            )
            .detach()
            .cpu()
            .numpy()[0, 0]
        )  ###试试重建不带上prompt部分
        audio_opt.append(audio)
        audio_opt.append(zero_wav)
        t4 = ttime()
    print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3))
    yield hps.data.sampling_rate, (np.concatenate(audio_opt, 0) * 32768).astype(
        np.int16
    )


splits = {
    ",",
    "。",
    "?",
    "!",
    ",",
    ".",
    "?",
    "!",
    "~",
    ":",
    ":",
    "—",
    "…",
}  # 不考虑省略号


def split(todo_text):
    todo_text = todo_text.replace("……", "。").replace("——", ",")
    if todo_text[-1] not in splits:
        todo_text += "。"
    i_split_head = i_split_tail = 0
    len_text = len(todo_text)
    todo_texts = []
    while 1:
        if i_split_head >= len_text:
            break  # 结尾一定有标点,所以直接跳出即可,最后一段在上次已加入
        if todo_text[i_split_head] in splits:
            i_split_head += 1
            todo_texts.append(todo_text[i_split_tail:i_split_head])
            i_split_tail = i_split_head
        else:
            i_split_head += 1
    return todo_texts


def cut1(inp):
    inp = inp.strip("\n")
    inps = split(inp)
    split_idx = list(range(0, len(inps), 5))
    split_idx[-1] = None
    if len(split_idx) > 1:
        opts = []
        for idx in range(len(split_idx) - 1):
            opts.append("".join(inps[split_idx[idx] : split_idx[idx + 1]]))
    else:
        opts = [inp]
    return "\n".join(opts)


def cut2(inp):
    inp = inp.strip("\n")
    inps = split(inp)
    if len(inps) < 2:
        return [inp]
    opts = []
    summ = 0
    tmp_str = ""
    for i in range(len(inps)):
        summ += len(inps[i])
        tmp_str += inps[i]
        if summ > 50:
            summ = 0
            opts.append(tmp_str)
            tmp_str = ""
    if tmp_str != "":
        opts.append(tmp_str)
    if len(opts[-1]) < 50:  ##如果最后一个太短了,和前一个合一起
        opts[-2] = opts[-2] + opts[-1]
        opts = opts[:-1]
    return "\n".join(opts)


def cut3(inp):
    inp = inp.strip("\n")
    return "\n".join(["%s。" % item for item in inp.strip("。").split("。")])


with gr.Blocks(title="GPT-SoVITS WebUI") as app:
    gr.Markdown(
        value="本软件以MIT协议开源, 作者不对软件具备任何控制力, 使用软件者、传播软件导出的声音者自负全责. <br>如不认可该条款, 则不能使用或引用软件包内任何代码和文件. 详见根目录<b>LICENSE</b>."
    )
    # with gr.Tabs():
    #     with gr.TabItem(i18n("伴奏人声分离&去混响&去回声")):
    with gr.Group():
        gr.Markdown(value="*请上传并填写参考信息")
        with gr.Row():
            inp_ref = gr.Audio(label="请上传参考音频", type="filepath")
            prompt_text = gr.Textbox(label="参考音频的文本", value="")
            prompt_language = gr.Dropdown(
                label="参考音频的语种", choices=["中文", "英文", "日文"], value="中文"
            )
        gr.Markdown(value="*请填写需要合成的目标文本")
        with gr.Row():
            text = gr.Textbox(label="需要合成的文本", value="")
            text_language = gr.Dropdown(
                label="需要合成的语种", choices=["中文", "英文", "日文"], value="中文"
            )
            inference_button = gr.Button("合成语音", variant="primary")
            output = gr.Audio(label="输出的语音")
        inference_button.click(
            get_tts_wav,
            [inp_ref, prompt_text, prompt_language, text, text_language],
            [output],
        )

        gr.Markdown(value="文本切分工具。太长的文本合成出来效果不一定好,所以太长建议先切。合成会根据文本的换行分开合成再拼起来。")
        with gr.Row():
            text_inp = gr.Textbox(label="需要合成的切分前文本", value="")
            button1 = gr.Button("凑五句一切", variant="primary")
            button2 = gr.Button("凑50字一切", variant="primary")
            button3 = gr.Button("按中文句号。切", variant="primary")
            text_opt = gr.Textbox(label="切分后文本", value="")
            button1.click(cut1, [text_inp], [text_opt])
            button2.click(cut2, [text_inp], [text_opt])
            button3.click(cut3, [text_inp], [text_opt])
        gr.Markdown(value="后续将支持混合语种编码文本输入。")

app.queue(concurrency_count=511, max_size=1022).launch(
    server_name="0.0.0.0",
    inbrowser=True,
    server_port=infer_ttswebui,
    quiet=True,
)