forked from Artrajz/vits-simple-api
-
Notifications
You must be signed in to change notification settings - Fork 0
/
voice.py
243 lines (209 loc) · 10.3 KB
/
voice.py
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
import os
import librosa
from scipy.io.wavfile import write
from mel_processing import spectrogram_torch
from text import text_to_sequence, _clean_text
from models import SynthesizerTrn
from utils import utils
import commons
import sys
import re
import numpy as np
import torch
# torch.set_num_threads(1) #设置torch线程为1,防止多任务推理时服务崩溃,但flask仍然会使用多线程
from torch import no_grad, LongTensor, inference_mode, FloatTensor
import uuid
from io import BytesIO
from graiax import silkcoder
from utils.nlp import cut, sentence_split
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"torch:{torch.__version__}", f"GPU_available:{torch.cuda.is_available()}")
print(f'device:{device} device.type:{device.type}')
class vits:
def __init__(self, model, config, model_=None):
self.mode_type = None
self.hps_ms = utils.get_hparams_from_file(config)
self.n_speakers = self.hps_ms.data.n_speakers if 'n_speakers' in self.hps_ms.data.keys() else 0
self.n_symbols = len(self.hps_ms.symbols) if 'symbols' in self.hps_ms.keys() else 0
self.speakers = self.hps_ms.speakers if 'speakers' in self.hps_ms.keys() else ['0']
self.use_f0 = self.hps_ms.data.use_f0 if 'use_f0' in self.hps_ms.data.keys() else False
self.emotion_embedding = self.hps_ms.data.emotion_embedding if 'emotion_embedding' in self.hps_ms.data.keys() else False
self.net_g_ms = SynthesizerTrn(
self.n_symbols,
self.hps_ms.data.filter_length // 2 + 1,
self.hps_ms.train.segment_size // self.hps_ms.data.hop_length,
n_speakers=self.n_speakers,
emotion_embedding=self.emotion_embedding,
**self.hps_ms.model)
_ = self.net_g_ms.eval()
if self.n_symbols != 0:
if not self.emotion_embedding:
self.mode_type = "vits"
else:
self.mode_type = "w2v2"
else:
self.mode_type = "hubert-soft"
# load model
self.load_model(model, model_)
def load_model(self, model, model_=None):
utils.load_checkpoint(model, self.net_g_ms)
self.net_g_ms.to(device)
if self.mode_type == "hubert-soft":
from hubert_model import hubert_soft
self.hubert = hubert_soft(model_)
if self.mode_type == "w2v2":
import audonnx
self.w2v2 = audonnx.load(model_)
def get_cleaned_text(self, text, hps, cleaned=False):
if cleaned:
text_norm = text_to_sequence(text, hps.symbols, [])
else:
text_norm = text_to_sequence(text, hps.symbols, hps.data.text_cleaners)
if hps.data.add_blank:
text_norm = commons.intersperse(text_norm, 0)
text_norm = LongTensor(text_norm)
return text_norm
def get_label_value(self, label, default, warning_name='value', text=""):
value = re.search(rf'\[{label}=(.+?)\]', text)
if value:
try:
text = re.sub(rf'\[{label}=(.+?)\]', '', text, 1)
value = float(value.group(1))
except:
print(f'Invalid {warning_name}!')
sys.exit(1)
else:
value = default
if text == "":
return value
else:
return value, text
def get_label(self, text, label):
if f'[{label}]' in text:
return True, text.replace(f'[{label}]', '')
else:
return False, text
def return_speakers(self, escape=False):
return self.speakers
def encode(self, sampling_rate, audio, format):
with BytesIO() as f:
write(f, sampling_rate, audio)
if format == 'ogg':
with BytesIO() as o:
utils.wav2ogg(f, o)
return BytesIO(o.getvalue())
elif format == 'silk':
return BytesIO(silkcoder.encode(f))
elif format == 'wav':
return BytesIO(f.getvalue())
def infer(self, params):
try:
emotion = params.get("emotion")
except:
emotion = None
with no_grad():
x_tst = params.get("stn_tst").unsqueeze(0)
x_tst_lengths = LongTensor([params.get("stn_tst").size(0)])
audio = self.net_g_ms.infer(x_tst.to(device), x_tst_lengths.to(device), sid=params.get("sid").to(device),
noise_scale=params.get("noise_scale"),
noise_scale_w=params.get("noise_scale_w"),
length_scale=params.get("length_scale"),
emotion_embedding=emotion)[0][0, 0].data.float().cpu().numpy()
torch.cuda.empty_cache()
return audio
def get_infer_param(self, length, noise, noisew, text=None, speaker_id=None, target_id=None, audio_path=None,
emotion=None):
emotion = None
if self.mode_type != "hubert-soft":
length_scale, text = self.get_label_value('LENGTH', length, 'length scale', text)
noise_scale, text = self.get_label_value('NOISE', noise, 'noise scale', text)
noise_scale_w, text = self.get_label_value('NOISEW', noisew, 'deviation of noise', text)
cleaned, text = self.get_label(text, 'CLEANED')
stn_tst = self.get_cleaned_text(text, self.hps_ms, cleaned=cleaned)
sid = LongTensor([speaker_id])
if self.mode_type == "w2v2":
emotion_reference = input('Path of an emotion reference: ')
if emotion_reference.endswith('.npy'):
emotion = np.load(emotion_reference)
emotion = FloatTensor(emotion).unsqueeze(0)
else:
audio16000, sampling_rate = librosa.load(
emotion_reference, sr=16000, mono=True)
emotion = self.w2v2(audio16000, sampling_rate)[
'hidden_states']
emotion_reference = re.sub(
r'\..*$', '', emotion_reference)
np.save(emotion_reference, emotion.squeeze(0))
emotion = FloatTensor(emotion)
elif self.mode_type == "hubert-soft":
if self.use_f0:
audio, sampling_rate = librosa.load(
audio_path, sr=self.hps_ms.data.sampling_rate, mono=True)
audio16000 = librosa.resample(
audio, orig_sr=sampling_rate, target_sr=16000)
else:
audio16000, sampling_rate = librosa.load(
audio_path, sr=16000, mono=True)
length_scale = self.get_label_value('LENGTH', length, 'length scale')
noise_scale = self.get_label_value('NOISE', noise, 'noise scale')
noise_scale_w = self.get_label_value('NOISEW', noisew, 'deviation of noise')
with inference_mode():
units = self.hubert.units(FloatTensor(audio16000).unsqueeze(0).unsqueeze(0)).squeeze(0).numpy()
if self.use_f0:
f0_scale = self.get_label_value('F0', 1, 'f0 scale')
f0 = librosa.pyin(audio,
sr=sampling_rate,
fmin=librosa.note_to_hz('C0'),
fmax=librosa.note_to_hz('C7'),
frame_length=1780)[0]
target_length = len(units[:, 0])
f0 = np.nan_to_num(np.interp(np.arange(0, len(f0) * target_length, len(f0)) / target_length,
np.arange(0, len(f0)), f0)) * f0_scale
units[:, 0] = f0 / 10
stn_tst = FloatTensor(units)
sid = LongTensor([target_id])
params = {"length_scale": length_scale, "noise_scale": noise_scale,
"noise_scale_w": noise_scale_w, "stn_tst": stn_tst,
"sid": sid, "emotion": emotion}
return params
def create_infer_task(self, text=None, speaker_id=None, format=None, length=1, noise=0.667, noisew=0.8,
target_id=None, audio_path=None, max=50, lang="auto"):
# params = self.get_infer_param(text=text, speaker_id=speaker_id, length=length, noise=noise, noisew=noisew,
# target_id=target_id)
tasks = []
if self.mode_type == "vits":
sentence_list = sentence_split(text, max, lang)
for sentence in sentence_list:
tasks.append(
self.get_infer_param(text=sentence, speaker_id=speaker_id, length=length, noise=noise,
noisew=noisew,
target_id=target_id))
audios = []
for task in tasks:
audios.append(self.infer(task))
audio = np.concatenate(audios, axis=0)
elif self.mode_type == "hubert-soft":
params = self.get_infer_param(speaker_id=speaker_id, length=length, noise=noise, noisew=noisew,
target_id=target_id, audio_path=audio_path)
audio = self.infer(params)
return self.encode(self.hps_ms.data.sampling_rate, audio, format)
def voice_conversion(self, audio_path, original_id, target_id):
audio = utils.load_audio_to_torch(
audio_path, self.hps_ms.data.sampling_rate)
y = audio.unsqueeze(0)
spec = spectrogram_torch(y, self.hps_ms.data.filter_length,
self.hps_ms.data.sampling_rate, self.hps_ms.data.hop_length,
self.hps_ms.data.win_length,
center=False)
spec_lengths = LongTensor([spec.size(-1)])
sid_src = LongTensor([original_id])
with no_grad():
sid_tgt = LongTensor([target_id])
audio = self.net_g_ms.voice_conversion(spec.to(device),
spec_lengths.to(device),
sid_src=sid_src.to(device),
sid_tgt=sid_tgt.to(device))[0][0, 0].data.cpu().float().numpy()
torch.cuda.empty_cache()
with BytesIO() as f:
write(f, self.hps_ms.data.sampling_rate, audio)
return BytesIO(f.getvalue())