Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,384 +1,384 @@
|
|
1 |
-
import os
|
2 |
-
import sys
|
3 |
-
# to avoid the modified user.pth file
|
4 |
-
cnhubert_base_path = "GPT_SoVITS\pretrained_models\chinese-hubert-base"
|
5 |
-
bert_path = "GPT_SoVITS\pretrained_models\chinese-roberta-wwm-ext-large"
|
6 |
-
os.environ["version"] = 'v2'
|
7 |
-
now_dir = os.getcwd()
|
8 |
-
sys.path.insert(0, now_dir)
|
9 |
-
import gradio as gr
|
10 |
-
from transformers import AutoModelForMaskedLM, AutoTokenizer
|
11 |
-
import numpy as np
|
12 |
-
from pathlib import Path
|
13 |
-
import os,librosa,torch
|
14 |
-
from scipy.io.wavfile import write as wavwrite
|
15 |
-
from GPT_SoVITS.feature_extractor import cnhubert
|
16 |
-
cnhubert.cnhubert_base_path=cnhubert_base_path
|
17 |
-
from GPT_SoVITS.module.models import SynthesizerTrn
|
18 |
-
from GPT_SoVITS.AR.models.t2s_lightning_module import Text2SemanticLightningModule
|
19 |
-
from GPT_SoVITS.text import cleaned_text_to_sequence
|
20 |
-
from GPT_SoVITS.text.cleaner import clean_text
|
21 |
-
from time import time as ttime
|
22 |
-
from GPT_SoVITS.module.mel_processing import spectrogram_torch
|
23 |
-
import tempfile
|
24 |
-
from tools.my_utils import load_audio
|
25 |
-
import os
|
26 |
-
import json
|
27 |
-
|
28 |
-
################ End strange import and user.pth modification ################
|
29 |
-
|
30 |
-
# import pyopenjtalk
|
31 |
-
# cwd = os.getcwd()
|
32 |
-
# if os.path.exists(os.path.join(cwd,'user.dic')):
|
33 |
-
# pyopenjtalk.update_global_jtalk_with_user_dict(os.path.join(cwd, 'user.dic'))
|
34 |
-
|
35 |
-
|
36 |
-
import logging
|
37 |
-
logging.getLogger('httpx').setLevel(logging.WARNING)
|
38 |
-
logging.getLogger('httpcore').setLevel(logging.WARNING)
|
39 |
-
logging.getLogger('multipart').setLevel(logging.WARNING)
|
40 |
-
|
41 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
42 |
-
#device = "cpu"
|
43 |
-
is_half = False
|
44 |
-
|
45 |
-
tokenizer = AutoTokenizer.from_pretrained(bert_path)
|
46 |
-
bert_model=AutoModelForMaskedLM.from_pretrained(bert_path)
|
47 |
-
if(is_half==True):bert_model=bert_model.half().to(device)
|
48 |
-
else:bert_model=bert_model.to(device)
|
49 |
-
# bert_model=bert_model.to(device)
|
50 |
-
def get_bert_feature(text, word2ph): # Bert(不是HuBERT的特征计算)
|
51 |
-
with torch.no_grad():
|
52 |
-
inputs = tokenizer(text, return_tensors="pt")
|
53 |
-
for i in inputs:
|
54 |
-
inputs[i] = inputs[i].to(device)#####输入是long不用管精度问题,精度随bert_model
|
55 |
-
res = bert_model(**inputs, output_hidden_states=True)
|
56 |
-
res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()[1:-1]
|
57 |
-
assert len(word2ph) == len(text)
|
58 |
-
phone_level_feature = []
|
59 |
-
for i in range(len(word2ph)):
|
60 |
-
repeat_feature = res[i].repeat(word2ph[i], 1)
|
61 |
-
phone_level_feature.append(repeat_feature)
|
62 |
-
phone_level_feature = torch.cat(phone_level_feature, dim=0)
|
63 |
-
# if(is_half==True):phone_level_feature=phone_level_feature.half()
|
64 |
-
return phone_level_feature.T
|
65 |
-
|
66 |
-
loaded_sovits_model = [] # [(path, dict, model)]
|
67 |
-
loaded_gpt_model = []
|
68 |
-
ssl_model = cnhubert.get_model()
|
69 |
-
if (is_half == True):
|
70 |
-
ssl_model = ssl_model.half().to(device)
|
71 |
-
else:
|
72 |
-
ssl_model = ssl_model.to(device)
|
73 |
-
|
74 |
-
|
75 |
-
def load_model(sovits_path, gpt_path):
|
76 |
-
global ssl_model
|
77 |
-
global loaded_sovits_model
|
78 |
-
global loaded_gpt_model
|
79 |
-
vq_model = None
|
80 |
-
t2s_model = None
|
81 |
-
dict_s2 = None
|
82 |
-
dict_s1 = None
|
83 |
-
hps = None
|
84 |
-
for path, dict_s2_, model in loaded_sovits_model:
|
85 |
-
if path == sovits_path:
|
86 |
-
vq_model = model
|
87 |
-
dict_s2 = dict_s2_
|
88 |
-
break
|
89 |
-
for path, dict_s1_, model in loaded_gpt_model:
|
90 |
-
if path == gpt_path:
|
91 |
-
t2s_model = model
|
92 |
-
dict_s1 = dict_s1_
|
93 |
-
break
|
94 |
-
|
95 |
-
if dict_s2 is None:
|
96 |
-
dict_s2 = torch.load(sovits_path, map_location="cpu")
|
97 |
-
hps = dict_s2["config"]
|
98 |
-
|
99 |
-
if dict_s1 is None:
|
100 |
-
dict_s1 = torch.load(gpt_path, map_location="cpu")
|
101 |
-
config = dict_s1["config"]
|
102 |
-
class DictToAttrRecursive:
|
103 |
-
def __init__(self, input_dict):
|
104 |
-
for key, value in input_dict.items():
|
105 |
-
if isinstance(value, dict):
|
106 |
-
# 如果值是字典,递归调用构造函数
|
107 |
-
setattr(self, key, DictToAttrRecursive(value))
|
108 |
-
else:
|
109 |
-
setattr(self, key, value)
|
110 |
-
|
111 |
-
hps = DictToAttrRecursive(hps)
|
112 |
-
hps.model.semantic_frame_rate = "25hz"
|
113 |
-
|
114 |
-
|
115 |
-
if not vq_model:
|
116 |
-
vq_model = SynthesizerTrn(
|
117 |
-
hps.data.filter_length // 2 + 1,
|
118 |
-
hps.train.segment_size // hps.data.hop_length,
|
119 |
-
n_speakers=hps.data.n_speakers,
|
120 |
-
**hps.model)
|
121 |
-
if (is_half == True):
|
122 |
-
vq_model = vq_model.half().to(device)
|
123 |
-
else:
|
124 |
-
vq_model = vq_model.to(device)
|
125 |
-
vq_model.eval()
|
126 |
-
vq_model.load_state_dict(dict_s2["weight"], strict=False)
|
127 |
-
loaded_sovits_model.append((sovits_path, dict_s2, vq_model))
|
128 |
-
hz = 50
|
129 |
-
max_sec = config['data']['max_sec']
|
130 |
-
if not t2s_model:
|
131 |
-
t2s_model = Text2SemanticLightningModule(config, "ojbk", is_train=False)
|
132 |
-
t2s_model.load_state_dict(dict_s1["weight"])
|
133 |
-
if (is_half == True): t2s_model = t2s_model.half()
|
134 |
-
t2s_model = t2s_model.to(device)
|
135 |
-
t2s_model.eval()
|
136 |
-
total = sum([param.nelement() for param in t2s_model.parameters()])
|
137 |
-
loaded_gpt_model.append((gpt_path, dict_s1, t2s_model))
|
138 |
-
return vq_model, ssl_model, t2s_model, hps, config, hz, max_sec
|
139 |
-
|
140 |
-
|
141 |
-
def get_spepc(hps, filename):
|
142 |
-
audio=load_audio(filename,int(hps.data.sampling_rate))
|
143 |
-
audio = audio / np.max(np.abs(audio))
|
144 |
-
audio=torch.FloatTensor(audio)
|
145 |
-
audio_norm = audio
|
146 |
-
# audio_norm = audio / torch.max(torch.abs(audio))
|
147 |
-
audio_norm = audio_norm.unsqueeze(0)
|
148 |
-
spec = spectrogram_torch(audio_norm, hps.data.filter_length,hps.data.sampling_rate, hps.data.hop_length, hps.data.win_length,center=False)
|
149 |
-
return spec
|
150 |
-
|
151 |
-
def create_tts_fn(vq_model, ssl_model, t2s_model, hps, config, hz, max_sec):
|
152 |
-
def tts_fn(ref_wav_path, prompt_text, prompt_language, target_phone, text_language, target_text = None):
|
153 |
-
t0 = ttime()
|
154 |
-
prompt_text=prompt_text.strip()
|
155 |
-
prompt_language=prompt_language
|
156 |
-
with torch.no_grad():
|
157 |
-
wav16k, sr = librosa.load(ref_wav_path, sr=16000, mono=False)
|
158 |
-
direction = np.array([1,1])
|
159 |
-
if wav16k.ndim == 2:
|
160 |
-
power = np.sum(np.abs(wav16k) ** 2, axis=1)
|
161 |
-
direction = power / np.sum(power)
|
162 |
-
wav16k = (wav16k[0] + wav16k[1]) / 2
|
163 |
-
#
|
164 |
-
# maxx=0.95
|
165 |
-
# tmp_max = np.abs(wav16k).max()
|
166 |
-
# alpha=0.5
|
167 |
-
# wav16k = (wav16k / tmp_max * (maxx * alpha*32768)) + ((1 - alpha)*32768) * wav16k
|
168 |
-
#在这里归一化
|
169 |
-
#print(max(np.abs(wav16k)))
|
170 |
-
#wav16k = wav16k / np.max(np.abs(wav16k))
|
171 |
-
#print(max(np.abs(wav16k)))
|
172 |
-
# 添加0.3s的静音
|
173 |
-
wav16k = np.concatenate([wav16k, np.zeros(int(hps.data.sampling_rate * 0.3)),])
|
174 |
-
wav16k = torch.from_numpy(wav16k)
|
175 |
-
wav16k = wav16k.float()
|
176 |
-
if(is_half==True):wav16k=wav16k.half().to(device)
|
177 |
-
else:wav16k=wav16k.to(device)
|
178 |
-
ssl_content = ssl_model.model(wav16k.unsqueeze(0))["last_hidden_state"].transpose(1, 2)#.float()
|
179 |
-
codes = vq_model.extract_latent(ssl_content)
|
180 |
-
prompt_semantic = codes[0, 0]
|
181 |
-
t1 = ttime()
|
182 |
-
phones1, word2ph1, norm_text1 = clean_text(prompt_text, prompt_language)
|
183 |
-
phones1=cleaned_text_to_sequence(phones1)
|
184 |
-
#texts=text.split("\n")
|
185 |
-
audio_opt = []
|
186 |
-
zero_wav=np.zeros((2, int(hps.data.sampling_rate*0.3)),dtype=np.float16 if is_half==True else np.float32)
|
187 |
-
phones = get_phone_from_str_list(target_phone, text_language)
|
188 |
-
for phones2 in phones:
|
189 |
-
if(len(phones2) == 0):
|
190 |
-
continue
|
191 |
-
if(len(phones2) == 1 and phones2[0] == ""):
|
192 |
-
continue
|
193 |
-
#phones2, word2ph2, norm_text2 = clean_text(text, text_language)
|
194 |
-
phones2 = cleaned_text_to_sequence(phones2)
|
195 |
-
#if(prompt_language=="zh"):bert1 = get_bert_feature(norm_text1, word2ph1).to(device)
|
196 |
-
bert1 = torch.zeros((1024, len(phones1)),dtype=torch.float16 if is_half==True else torch.float32).to(device)
|
197 |
-
#if(text_language=="zh"):bert2 = get_bert_feature(norm_text2, word2ph2).to(device)
|
198 |
-
bert2 = torch.zeros((1024, len(phones2))).to(bert1)
|
199 |
-
bert = torch.cat([bert1, bert2], 1)
|
200 |
-
|
201 |
-
all_phoneme_ids = torch.LongTensor(phones1+phones2).to(device).unsqueeze(0)
|
202 |
-
bert = bert.to(device).unsqueeze(0)
|
203 |
-
all_phoneme_len = torch.tensor([all_phoneme_ids.shape[-1]]).to(device)
|
204 |
-
prompt = prompt_semantic.unsqueeze(0).to(device)
|
205 |
-
t2 = ttime()
|
206 |
-
idx = 0
|
207 |
-
cnt = 0
|
208 |
-
while idx == 0 and cnt < 2:
|
209 |
-
with torch.no_grad():
|
210 |
-
# pred_semantic = t2s_model.model.infer
|
211 |
-
pred_semantic,idx = t2s_model.model.infer_panel(
|
212 |
-
all_phoneme_ids,
|
213 |
-
all_phoneme_len,
|
214 |
-
prompt,
|
215 |
-
bert,
|
216 |
-
# prompt_phone_len=ph_offset,
|
217 |
-
top_k=config['inference']['top_k'],
|
218 |
-
early_stop_num=hz * max_sec)
|
219 |
-
t3 = ttime()
|
220 |
-
cnt+=1
|
221 |
-
if idx == 0:
|
222 |
-
return "Error: Generation failure: bad zero prediction.", None
|
223 |
-
pred_semantic = pred_semantic[:,-idx:].unsqueeze(0) # .unsqueeze(0)#mq要多unsqueeze一次
|
224 |
-
refer = get_spepc(hps, ref_wav_path)#.to(device)
|
225 |
-
if(is_half==True):refer=refer.half().to(device)
|
226 |
-
else:refer=refer.to(device)
|
227 |
-
# audio = vq_model.decode(pred_semantic, all_phoneme_ids, refer).detach().cpu().numpy()[0, 0]
|
228 |
-
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
|
229 |
-
# direction乘上,变双通道
|
230 |
-
# 强制0.5
|
231 |
-
direction = np.array([1, 1])
|
232 |
-
audio = np.expand_dims(audio, 0) * direction[:, np.newaxis]
|
233 |
-
audio_opt.append(audio)
|
234 |
-
audio_opt.append(zero_wav)
|
235 |
-
t4 = ttime()
|
236 |
-
|
237 |
-
audio = (hps.data.sampling_rate,(np.concatenate(audio_opt, axis=1)*32768).astype(np.int16).T)
|
238 |
-
prefix_1 = prompt_text[:8].replace(" ", "_").replace("\n", "_").replace("?","_").replace("!","_").replace(",","_")
|
239 |
-
prefix_2 = target_text[:8].replace(" ", "_").replace("\n", "_").replace("?","_").replace("!","_").replace(",","_")
|
240 |
-
filename = tempfile.mktemp(suffix=".wav",prefix=f"{prefix_1}_{prefix_2}_")
|
241 |
-
#audiosegment.from_numpy_array(audio[1].T, framerate=audio[0]).export(filename, format="WAV")
|
242 |
-
wavwrite(filename, audio[0], audio[1])
|
243 |
-
return "Success", audio, filename
|
244 |
-
return tts_fn
|
245 |
-
|
246 |
-
|
247 |
-
def get_str_list_from_phone(text, text_language):
|
248 |
-
# raw文本过g2p得到音素列表,再转成字符串
|
249 |
-
# 注意,这里的text是一个段落,可能包含多个句子
|
250 |
-
# 段落间\n分割,音素间空格分割
|
251 |
-
print(text)
|
252 |
-
texts=text.split("\n")
|
253 |
-
phone_list = []
|
254 |
-
for text in texts:
|
255 |
-
phones2, word2ph2, norm_text2 = clean_text(text, text_language)
|
256 |
-
phone_list.append(" ".join(phones2))
|
257 |
-
return "\n".join(phone_list)
|
258 |
-
|
259 |
-
def get_phone_from_str_list(str_list:str, language:str = 'ja'):
|
260 |
-
# 从音素字符串中得到音素列表
|
261 |
-
# 注意,这里的text是一个段落,可能包含多个句子
|
262 |
-
# 段落间\n分割,音素间空格分割
|
263 |
-
sentences = str_list.split("\n")
|
264 |
-
phones = []
|
265 |
-
for sentence in sentences:
|
266 |
-
phones.append(sentence.split(" "))
|
267 |
-
return phones
|
268 |
-
|
269 |
-
splits={",","。","?","!",",",".","?","!","~",":",":","—","…",}#不考虑省略号
|
270 |
-
def split(todo_text):
|
271 |
-
todo_text = todo_text.replace("……", "。").replace("——", ",")
|
272 |
-
if (todo_text[-1] not in splits): todo_text += "。"
|
273 |
-
i_split_head = i_split_tail = 0
|
274 |
-
len_text = len(todo_text)
|
275 |
-
todo_texts = []
|
276 |
-
while (1):
|
277 |
-
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
|
278 |
-
if (todo_text[i_split_head] in splits):
|
279 |
-
i_split_head += 1
|
280 |
-
todo_texts.append(todo_text[i_split_tail:i_split_head])
|
281 |
-
i_split_tail = i_split_head
|
282 |
-
else:
|
283 |
-
i_split_head += 1
|
284 |
-
return todo_texts
|
285 |
-
|
286 |
-
|
287 |
-
def change_reference_audio(prompt_text, transcripts):
|
288 |
-
return transcripts[prompt_text]
|
289 |
-
|
290 |
-
|
291 |
-
models = []
|
292 |
-
models_info = json.load(open("./models/models_info.json", "r", encoding="utf-8"))
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
for i, info in models_info.items():
|
297 |
-
title = info['title']
|
298 |
-
cover = info['cover']
|
299 |
-
gpt_weight = info['gpt_weight']
|
300 |
-
sovits_weight = info['sovits_weight']
|
301 |
-
example_reference = info['example_reference']
|
302 |
-
transcripts = {}
|
303 |
-
transcript_path = info["transcript_path"]
|
304 |
-
path = os.path.dirname(transcript_path)
|
305 |
-
with open(transcript_path, 'r', encoding='utf-8') as file:
|
306 |
-
for line in file:
|
307 |
-
line = line.strip().replace("\\", "/")
|
308 |
-
items = line.split("|")
|
309 |
-
wav,t = items[0], items[-1]
|
310 |
-
wav = os.path.basename(wav)
|
311 |
-
transcripts[t] = os.path.join(os.path.join(path,"reference_audio"), wav)
|
312 |
-
|
313 |
-
vq_model, ssl_model, t2s_model, hps, config, hz, max_sec = load_model(sovits_weight, gpt_weight)
|
314 |
-
|
315 |
-
|
316 |
-
models.append(
|
317 |
-
(
|
318 |
-
i,
|
319 |
-
title,
|
320 |
-
cover,
|
321 |
-
transcripts,
|
322 |
-
example_reference,
|
323 |
-
create_tts_fn(
|
324 |
-
vq_model, ssl_model, t2s_model, hps, config, hz, max_sec
|
325 |
-
)
|
326 |
-
)
|
327 |
-
)
|
328 |
-
with gr.Blocks() as app:
|
329 |
-
gr.Markdown(
|
330 |
-
"# <center> GPT-SoVITS Demo\n"
|
331 |
-
)
|
332 |
-
with gr.Tabs():
|
333 |
-
for (name, title, cover, transcripts, example_reference, tts_fn) in models:
|
334 |
-
with gr.TabItem(name):
|
335 |
-
with gr.Row():
|
336 |
-
gr.Markdown(
|
337 |
-
'<div align="center">'
|
338 |
-
f'<a><strong>{title}</strong></a>'
|
339 |
-
'</div>')
|
340 |
-
with gr.Row():
|
341 |
-
with gr.Column():
|
342 |
-
prompt_text = gr.Dropdown(
|
343 |
-
label="Transcript of the Reference Audio",
|
344 |
-
value=example_reference if example_reference in transcripts else list(transcripts.keys())[0],
|
345 |
-
choices=list(transcripts.keys())
|
346 |
-
)
|
347 |
-
inp_ref_audio = gr.Audio(
|
348 |
-
label="Reference Audio",
|
349 |
-
type="filepath",
|
350 |
-
interactive=False,
|
351 |
-
value=transcripts[example_reference] if example_reference in transcripts else list(transcripts.values())[0]
|
352 |
-
)
|
353 |
-
transcripts_state = gr.State(value=transcripts)
|
354 |
-
prompt_text.change(
|
355 |
-
fn=change_reference_audio,
|
356 |
-
inputs=[prompt_text, transcripts_state],
|
357 |
-
outputs=[inp_ref_audio]
|
358 |
-
)
|
359 |
-
prompt_language = gr.State(value="ja")
|
360 |
-
with gr.Column():
|
361 |
-
text = gr.Textbox(label="Input Text", value="私はお兄ちゃんのだいだいだーいすきな妹なんだから、言うことなんでも聞いてくれますよね!")
|
362 |
-
text_language = gr.Dropdown(
|
363 |
-
label="Language",
|
364 |
-
choices=["ja"],
|
365 |
-
value="ja"
|
366 |
-
)
|
367 |
-
clean_button = gr.Button("Clean Text", variant="primary")
|
368 |
-
inference_button = gr.Button("Generate", variant="primary")
|
369 |
-
cleaned_text = gr.Textbox(label="Cleaned Text")
|
370 |
-
output = gr.Audio(label="Output Audio")
|
371 |
-
output_file = gr.File(label="Output Audio File")
|
372 |
-
om = gr.Textbox(label="Output Message")
|
373 |
-
clean_button.click(
|
374 |
-
fn=get_str_list_from_phone,
|
375 |
-
inputs=[text, text_language],
|
376 |
-
outputs=[cleaned_text]
|
377 |
-
)
|
378 |
-
inference_button.click(
|
379 |
-
fn=tts_fn,
|
380 |
-
inputs=[inp_ref_audio, prompt_text, prompt_language, cleaned_text, text_language, text],
|
381 |
-
outputs=[om, output, output_file]
|
382 |
-
)
|
383 |
-
|
384 |
app.launch(share=True)
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
# to avoid the modified user.pth file
|
4 |
+
cnhubert_base_path = "GPT_SoVITS\pretrained_models\chinese-hubert-base"
|
5 |
+
bert_path = "GPT_SoVITS\pretrained_models\chinese-roberta-wwm-ext-large"
|
6 |
+
os.environ["version"] = 'v2'
|
7 |
+
now_dir = os.getcwd()
|
8 |
+
sys.path.insert(0, now_dir)
|
9 |
+
import gradio as gr
|
10 |
+
from transformers import AutoModelForMaskedLM, AutoTokenizer
|
11 |
+
import numpy as np
|
12 |
+
from pathlib import Path
|
13 |
+
import os,librosa,torch
|
14 |
+
from scipy.io.wavfile import write as wavwrite
|
15 |
+
from GPT_SoVITS.feature_extractor import cnhubert
|
16 |
+
cnhubert.cnhubert_base_path=cnhubert_base_path
|
17 |
+
from GPT_SoVITS.module.models import SynthesizerTrn
|
18 |
+
from GPT_SoVITS.AR.models.t2s_lightning_module import Text2SemanticLightningModule
|
19 |
+
from GPT_SoVITS.text import cleaned_text_to_sequence
|
20 |
+
from GPT_SoVITS.text.cleaner import clean_text
|
21 |
+
from time import time as ttime
|
22 |
+
from GPT_SoVITS.module.mel_processing import spectrogram_torch
|
23 |
+
import tempfile
|
24 |
+
from tools.my_utils import load_audio
|
25 |
+
import os
|
26 |
+
import json
|
27 |
+
|
28 |
+
################ End strange import and user.pth modification ################
|
29 |
+
|
30 |
+
# import pyopenjtalk
|
31 |
+
# cwd = os.getcwd()
|
32 |
+
# if os.path.exists(os.path.join(cwd,'user.dic')):
|
33 |
+
# pyopenjtalk.update_global_jtalk_with_user_dict(os.path.join(cwd, 'user.dic'))
|
34 |
+
|
35 |
+
|
36 |
+
import logging
|
37 |
+
logging.getLogger('httpx').setLevel(logging.WARNING)
|
38 |
+
logging.getLogger('httpcore').setLevel(logging.WARNING)
|
39 |
+
logging.getLogger('multipart').setLevel(logging.WARNING)
|
40 |
+
|
41 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
42 |
+
#device = "cpu"
|
43 |
+
is_half = False
|
44 |
+
|
45 |
+
tokenizer = AutoTokenizer.from_pretrained(bert_path)
|
46 |
+
bert_model=AutoModelForMaskedLM.from_pretrained(bert_path)
|
47 |
+
if(is_half==True):bert_model=bert_model.half().to(device)
|
48 |
+
else:bert_model=bert_model.to(device)
|
49 |
+
# bert_model=bert_model.to(device)
|
50 |
+
def get_bert_feature(text, word2ph): # Bert(不是HuBERT的特征计算)
|
51 |
+
with torch.no_grad():
|
52 |
+
inputs = tokenizer(text, return_tensors="pt")
|
53 |
+
for i in inputs:
|
54 |
+
inputs[i] = inputs[i].to(device)#####输入是long不用管精度问题,精度随bert_model
|
55 |
+
res = bert_model(**inputs, output_hidden_states=True)
|
56 |
+
res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()[1:-1]
|
57 |
+
assert len(word2ph) == len(text)
|
58 |
+
phone_level_feature = []
|
59 |
+
for i in range(len(word2ph)):
|
60 |
+
repeat_feature = res[i].repeat(word2ph[i], 1)
|
61 |
+
phone_level_feature.append(repeat_feature)
|
62 |
+
phone_level_feature = torch.cat(phone_level_feature, dim=0)
|
63 |
+
# if(is_half==True):phone_level_feature=phone_level_feature.half()
|
64 |
+
return phone_level_feature.T
|
65 |
+
|
66 |
+
loaded_sovits_model = [] # [(path, dict, model)]
|
67 |
+
loaded_gpt_model = []
|
68 |
+
ssl_model = cnhubert.get_model()
|
69 |
+
if (is_half == True):
|
70 |
+
ssl_model = ssl_model.half().to(device)
|
71 |
+
else:
|
72 |
+
ssl_model = ssl_model.to(device)
|
73 |
+
|
74 |
+
|
75 |
+
def load_model(sovits_path, gpt_path):
|
76 |
+
global ssl_model
|
77 |
+
global loaded_sovits_model
|
78 |
+
global loaded_gpt_model
|
79 |
+
vq_model = None
|
80 |
+
t2s_model = None
|
81 |
+
dict_s2 = None
|
82 |
+
dict_s1 = None
|
83 |
+
hps = None
|
84 |
+
for path, dict_s2_, model in loaded_sovits_model:
|
85 |
+
if path == sovits_path:
|
86 |
+
vq_model = model
|
87 |
+
dict_s2 = dict_s2_
|
88 |
+
break
|
89 |
+
for path, dict_s1_, model in loaded_gpt_model:
|
90 |
+
if path == gpt_path:
|
91 |
+
t2s_model = model
|
92 |
+
dict_s1 = dict_s1_
|
93 |
+
break
|
94 |
+
|
95 |
+
if dict_s2 is None:
|
96 |
+
dict_s2 = torch.load(sovits_path, map_location="cpu")
|
97 |
+
hps = dict_s2["config"]
|
98 |
+
|
99 |
+
if dict_s1 is None:
|
100 |
+
dict_s1 = torch.load(gpt_path, map_location="cpu")
|
101 |
+
config = dict_s1["config"]
|
102 |
+
class DictToAttrRecursive:
|
103 |
+
def __init__(self, input_dict):
|
104 |
+
for key, value in input_dict.items():
|
105 |
+
if isinstance(value, dict):
|
106 |
+
# 如果值是字典,递归调用构造函数
|
107 |
+
setattr(self, key, DictToAttrRecursive(value))
|
108 |
+
else:
|
109 |
+
setattr(self, key, value)
|
110 |
+
|
111 |
+
hps = DictToAttrRecursive(hps)
|
112 |
+
hps.model.semantic_frame_rate = "25hz"
|
113 |
+
|
114 |
+
|
115 |
+
if not vq_model:
|
116 |
+
vq_model = SynthesizerTrn(
|
117 |
+
hps.data.filter_length // 2 + 1,
|
118 |
+
hps.train.segment_size // hps.data.hop_length,
|
119 |
+
n_speakers=hps.data.n_speakers,
|
120 |
+
**hps.model)
|
121 |
+
if (is_half == True):
|
122 |
+
vq_model = vq_model.half().to(device)
|
123 |
+
else:
|
124 |
+
vq_model = vq_model.to(device)
|
125 |
+
vq_model.eval()
|
126 |
+
vq_model.load_state_dict(dict_s2["weight"], strict=False)
|
127 |
+
loaded_sovits_model.append((sovits_path, dict_s2, vq_model))
|
128 |
+
hz = 50
|
129 |
+
max_sec = config['data']['max_sec']
|
130 |
+
if not t2s_model:
|
131 |
+
t2s_model = Text2SemanticLightningModule(config, "ojbk", is_train=False)
|
132 |
+
t2s_model.load_state_dict(dict_s1["weight"])
|
133 |
+
if (is_half == True): t2s_model = t2s_model.half()
|
134 |
+
t2s_model = t2s_model.to(device)
|
135 |
+
t2s_model.eval()
|
136 |
+
total = sum([param.nelement() for param in t2s_model.parameters()])
|
137 |
+
loaded_gpt_model.append((gpt_path, dict_s1, t2s_model))
|
138 |
+
return vq_model, ssl_model, t2s_model, hps, config, hz, max_sec
|
139 |
+
|
140 |
+
|
141 |
+
def get_spepc(hps, filename):
|
142 |
+
audio=load_audio(filename,int(hps.data.sampling_rate))
|
143 |
+
audio = audio / np.max(np.abs(audio))
|
144 |
+
audio=torch.FloatTensor(audio)
|
145 |
+
audio_norm = audio
|
146 |
+
# audio_norm = audio / torch.max(torch.abs(audio))
|
147 |
+
audio_norm = audio_norm.unsqueeze(0)
|
148 |
+
spec = spectrogram_torch(audio_norm, hps.data.filter_length,hps.data.sampling_rate, hps.data.hop_length, hps.data.win_length,center=False)
|
149 |
+
return spec
|
150 |
+
|
151 |
+
def create_tts_fn(vq_model, ssl_model, t2s_model, hps, config, hz, max_sec):
|
152 |
+
def tts_fn(ref_wav_path, prompt_text, prompt_language, target_phone, text_language, target_text = None):
|
153 |
+
t0 = ttime()
|
154 |
+
prompt_text=prompt_text.strip()
|
155 |
+
prompt_language=prompt_language
|
156 |
+
with torch.no_grad():
|
157 |
+
wav16k, sr = librosa.load(ref_wav_path, sr=16000, mono=False)
|
158 |
+
direction = np.array([1,1])
|
159 |
+
if wav16k.ndim == 2:
|
160 |
+
power = np.sum(np.abs(wav16k) ** 2, axis=1)
|
161 |
+
direction = power / np.sum(power)
|
162 |
+
wav16k = (wav16k[0] + wav16k[1]) / 2
|
163 |
+
#
|
164 |
+
# maxx=0.95
|
165 |
+
# tmp_max = np.abs(wav16k).max()
|
166 |
+
# alpha=0.5
|
167 |
+
# wav16k = (wav16k / tmp_max * (maxx * alpha*32768)) + ((1 - alpha)*32768) * wav16k
|
168 |
+
#在这里归一化
|
169 |
+
#print(max(np.abs(wav16k)))
|
170 |
+
#wav16k = wav16k / np.max(np.abs(wav16k))
|
171 |
+
#print(max(np.abs(wav16k)))
|
172 |
+
# 添加0.3s的静音
|
173 |
+
wav16k = np.concatenate([wav16k, np.zeros(int(hps.data.sampling_rate * 0.3)),])
|
174 |
+
wav16k = torch.from_numpy(wav16k)
|
175 |
+
wav16k = wav16k.float()
|
176 |
+
if(is_half==True):wav16k=wav16k.half().to(device)
|
177 |
+
else:wav16k=wav16k.to(device)
|
178 |
+
ssl_content = ssl_model.model(wav16k.unsqueeze(0))["last_hidden_state"].transpose(1, 2)#.float()
|
179 |
+
codes = vq_model.extract_latent(ssl_content)
|
180 |
+
prompt_semantic = codes[0, 0]
|
181 |
+
t1 = ttime()
|
182 |
+
phones1, word2ph1, norm_text1 = clean_text(prompt_text, prompt_language)
|
183 |
+
phones1=cleaned_text_to_sequence(phones1)
|
184 |
+
#texts=text.split("\n")
|
185 |
+
audio_opt = []
|
186 |
+
zero_wav=np.zeros((2, int(hps.data.sampling_rate*0.3)),dtype=np.float16 if is_half==True else np.float32)
|
187 |
+
phones = get_phone_from_str_list(target_phone, text_language)
|
188 |
+
for phones2 in phones:
|
189 |
+
if(len(phones2) == 0):
|
190 |
+
continue
|
191 |
+
if(len(phones2) == 1 and phones2[0] == ""):
|
192 |
+
continue
|
193 |
+
#phones2, word2ph2, norm_text2 = clean_text(text, text_language)
|
194 |
+
phones2 = cleaned_text_to_sequence(phones2)
|
195 |
+
#if(prompt_language=="zh"):bert1 = get_bert_feature(norm_text1, word2ph1).to(device)
|
196 |
+
bert1 = torch.zeros((1024, len(phones1)),dtype=torch.float16 if is_half==True else torch.float32).to(device)
|
197 |
+
#if(text_language=="zh"):bert2 = get_bert_feature(norm_text2, word2ph2).to(device)
|
198 |
+
bert2 = torch.zeros((1024, len(phones2))).to(bert1)
|
199 |
+
bert = torch.cat([bert1, bert2], 1)
|
200 |
+
|
201 |
+
all_phoneme_ids = torch.LongTensor(phones1+phones2).to(device).unsqueeze(0)
|
202 |
+
bert = bert.to(device).unsqueeze(0)
|
203 |
+
all_phoneme_len = torch.tensor([all_phoneme_ids.shape[-1]]).to(device)
|
204 |
+
prompt = prompt_semantic.unsqueeze(0).to(device)
|
205 |
+
t2 = ttime()
|
206 |
+
idx = 0
|
207 |
+
cnt = 0
|
208 |
+
while idx == 0 and cnt < 2:
|
209 |
+
with torch.no_grad():
|
210 |
+
# pred_semantic = t2s_model.model.infer
|
211 |
+
pred_semantic,idx = t2s_model.model.infer_panel(
|
212 |
+
all_phoneme_ids,
|
213 |
+
all_phoneme_len,
|
214 |
+
prompt,
|
215 |
+
bert,
|
216 |
+
# prompt_phone_len=ph_offset,
|
217 |
+
top_k=config['inference']['top_k'],
|
218 |
+
early_stop_num=hz * max_sec)
|
219 |
+
t3 = ttime()
|
220 |
+
cnt+=1
|
221 |
+
if idx == 0:
|
222 |
+
return "Error: Generation failure: bad zero prediction.", None
|
223 |
+
pred_semantic = pred_semantic[:,-idx:].unsqueeze(0) # .unsqueeze(0)#mq要多unsqueeze一次
|
224 |
+
refer = get_spepc(hps, ref_wav_path)#.to(device)
|
225 |
+
if(is_half==True):refer=refer.half().to(device)
|
226 |
+
else:refer=refer.to(device)
|
227 |
+
# audio = vq_model.decode(pred_semantic, all_phoneme_ids, refer).detach().cpu().numpy()[0, 0]
|
228 |
+
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
|
229 |
+
# direction乘上,变双通道
|
230 |
+
# 强制0.5
|
231 |
+
direction = np.array([1, 1])
|
232 |
+
audio = np.expand_dims(audio, 0) * direction[:, np.newaxis]
|
233 |
+
audio_opt.append(audio)
|
234 |
+
audio_opt.append(zero_wav)
|
235 |
+
t4 = ttime()
|
236 |
+
|
237 |
+
audio = (hps.data.sampling_rate,(np.concatenate(audio_opt, axis=1)*32768).astype(np.int16).T)
|
238 |
+
prefix_1 = prompt_text[:8].replace(" ", "_").replace("\n", "_").replace("?","_").replace("!","_").replace(",","_")
|
239 |
+
prefix_2 = target_text[:8].replace(" ", "_").replace("\n", "_").replace("?","_").replace("!","_").replace(",","_")
|
240 |
+
filename = tempfile.mktemp(suffix=".wav",prefix=f"{prefix_1}_{prefix_2}_")
|
241 |
+
#audiosegment.from_numpy_array(audio[1].T, framerate=audio[0]).export(filename, format="WAV")
|
242 |
+
wavwrite(filename, audio[0], audio[1])
|
243 |
+
return "Success", audio, filename
|
244 |
+
return tts_fn
|
245 |
+
|
246 |
+
|
247 |
+
def get_str_list_from_phone(text, text_language):
|
248 |
+
# raw文本过g2p得到音素列表,再转成字符串
|
249 |
+
# 注意,这里的text是一个段落,可能包含多个句子
|
250 |
+
# 段落间\n分割,音素间空格分割
|
251 |
+
print(text)
|
252 |
+
texts=text.split("\n")
|
253 |
+
phone_list = []
|
254 |
+
for text in texts:
|
255 |
+
phones2, word2ph2, norm_text2 = clean_text(text, text_language)
|
256 |
+
phone_list.append(" ".join(phones2))
|
257 |
+
return "\n".join(phone_list)
|
258 |
+
|
259 |
+
def get_phone_from_str_list(str_list:str, language:str = 'ja'):
|
260 |
+
# 从音素字符串中得到音素列表
|
261 |
+
# 注意,这里的text是一个段落,可能包含多个句子
|
262 |
+
# 段落间\n分割,音素间空格分割
|
263 |
+
sentences = str_list.split("\n")
|
264 |
+
phones = []
|
265 |
+
for sentence in sentences:
|
266 |
+
phones.append(sentence.split(" "))
|
267 |
+
return phones
|
268 |
+
|
269 |
+
splits={",","。","?","!",",",".","?","!","~",":",":","—","…",}#不考虑省略号
|
270 |
+
def split(todo_text):
|
271 |
+
todo_text = todo_text.replace("……", "。").replace("——", ",")
|
272 |
+
if (todo_text[-1] not in splits): todo_text += "。"
|
273 |
+
i_split_head = i_split_tail = 0
|
274 |
+
len_text = len(todo_text)
|
275 |
+
todo_texts = []
|
276 |
+
while (1):
|
277 |
+
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
|
278 |
+
if (todo_text[i_split_head] in splits):
|
279 |
+
i_split_head += 1
|
280 |
+
todo_texts.append(todo_text[i_split_tail:i_split_head])
|
281 |
+
i_split_tail = i_split_head
|
282 |
+
else:
|
283 |
+
i_split_head += 1
|
284 |
+
return todo_texts
|
285 |
+
|
286 |
+
|
287 |
+
def change_reference_audio(prompt_text, transcripts):
|
288 |
+
return transcripts[prompt_text]
|
289 |
+
|
290 |
+
|
291 |
+
models = []
|
292 |
+
models_info = json.load(open("./models/models_info.json", "r", encoding="utf-8"))
|
293 |
+
|
294 |
+
|
295 |
+
|
296 |
+
for i, info in models_info.items():
|
297 |
+
title = info['title']
|
298 |
+
cover = info['cover']
|
299 |
+
gpt_weight = info['gpt_weight']
|
300 |
+
sovits_weight = info['sovits_weight']
|
301 |
+
example_reference = info['example_reference']
|
302 |
+
transcripts = {}
|
303 |
+
transcript_path = info["transcript_path"]
|
304 |
+
path = os.path.dirname(transcript_path)
|
305 |
+
with open(transcript_path, 'r', encoding='utf-8') as file:
|
306 |
+
for line in file:
|
307 |
+
line = line.strip().replace("\\", "/")
|
308 |
+
items = line.split("|")
|
309 |
+
wav,t = items[0], items[-1]
|
310 |
+
wav = os.path.basename(wav)
|
311 |
+
transcripts[t] = os.path.join(os.path.join(path,"reference_audio"), wav)
|
312 |
+
|
313 |
+
vq_model, ssl_model, t2s_model, hps, config, hz, max_sec = load_model(sovits_weight, gpt_weight)
|
314 |
+
|
315 |
+
|
316 |
+
models.append(
|
317 |
+
(
|
318 |
+
i,
|
319 |
+
title,
|
320 |
+
cover,
|
321 |
+
transcripts,
|
322 |
+
example_reference,
|
323 |
+
create_tts_fn(
|
324 |
+
vq_model, ssl_model, t2s_model, hps, config, hz, max_sec
|
325 |
+
)
|
326 |
+
)
|
327 |
+
)
|
328 |
+
with gr.Blocks() as app:
|
329 |
+
gr.Markdown(
|
330 |
+
"# <center> GPT-SoVITS Demo\n"
|
331 |
+
)
|
332 |
+
with gr.Tabs():
|
333 |
+
for (name, title, cover, transcripts, example_reference, tts_fn) in models:
|
334 |
+
with gr.TabItem(name):
|
335 |
+
with gr.Row():
|
336 |
+
gr.Markdown(
|
337 |
+
'<div align="center">'
|
338 |
+
f'<a><strong>{title}</strong></a>'
|
339 |
+
'</div>')
|
340 |
+
with gr.Row():
|
341 |
+
with gr.Column():
|
342 |
+
prompt_text = gr.Dropdown(
|
343 |
+
label="Transcript of the Reference Audio",
|
344 |
+
value=example_reference if example_reference in transcripts else list(transcripts.keys())[0],
|
345 |
+
choices=list(transcripts.keys())
|
346 |
+
)
|
347 |
+
inp_ref_audio = gr.Audio(
|
348 |
+
label="Reference Audio",
|
349 |
+
type="filepath",
|
350 |
+
interactive=False,
|
351 |
+
value=transcripts[example_reference] if example_reference in transcripts else list(transcripts.values())[0]
|
352 |
+
)
|
353 |
+
transcripts_state = gr.State(value=transcripts)
|
354 |
+
prompt_text.change(
|
355 |
+
fn=change_reference_audio,
|
356 |
+
inputs=[prompt_text, transcripts_state],
|
357 |
+
outputs=[inp_ref_audio]
|
358 |
+
)
|
359 |
+
prompt_language = gr.State(value="ja")
|
360 |
+
with gr.Column():
|
361 |
+
text = gr.Textbox(label="Input Text", value="私はお兄ちゃんのだいだいだーいすきな妹なんだから、言うことなんでも聞いてくれますよね!")
|
362 |
+
text_language = gr.Dropdown(
|
363 |
+
label="Language",
|
364 |
+
choices=["ja"],
|
365 |
+
value="ja"
|
366 |
+
)
|
367 |
+
clean_button = gr.Button("Clean Text", variant="primary")
|
368 |
+
inference_button = gr.Button("Generate", variant="primary")
|
369 |
+
cleaned_text = gr.Textbox(label="Cleaned Text")
|
370 |
+
output = gr.Audio(label="Output Audio")
|
371 |
+
output_file = gr.File(label="Output Audio File")
|
372 |
+
om = gr.Textbox(label="Output Message")
|
373 |
+
clean_button.click(
|
374 |
+
fn=get_str_list_from_phone,
|
375 |
+
inputs=[text, text_language],
|
376 |
+
outputs=[cleaned_text]
|
377 |
+
)
|
378 |
+
inference_button.click(
|
379 |
+
fn=tts_fn,
|
380 |
+
inputs=[inp_ref_audio, prompt_text, prompt_language, cleaned_text, text_language, text],
|
381 |
+
outputs=[om, output, output_file]
|
382 |
+
)
|
383 |
+
|
384 |
app.launch(share=True)
|