-
Notifications
You must be signed in to change notification settings - Fork 368
/
utils_vad.py
597 lines (508 loc) · 21.2 KB
/
utils_vad.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
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
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
import torch
import torchaudio
from typing import List
from itertools import repeat
from collections import deque
import torch.nn.functional as F
torchaudio.set_audio_backend("soundfile") # switch backend
languages = ['ru', 'en', 'de', 'es']
class IterativeMedianMeter():
def __init__(self):
self.reset()
def reset(self):
self.median = 0
self.counts = {}
for i in range(0, 101, 1):
self.counts[i / 100] = 0
self.total_values = 0
def __call__(self, val):
self.total_values += 1
rounded = round(abs(val), 2)
self.counts[rounded] += 1
bin_sum = 0
for j in self.counts:
bin_sum += self.counts[j]
if bin_sum >= self.total_values / 2:
self.median = j
break
return self.median
def validate(model,
inputs: torch.Tensor):
with torch.no_grad():
outs = model(inputs)
return outs
def read_audio(path: str,
target_sr: int = 16000):
assert torchaudio.get_audio_backend() == 'soundfile'
wav, sr = torchaudio.load(path)
if wav.size(0) > 1:
wav = wav.mean(dim=0, keepdim=True)
if sr != target_sr:
transform = torchaudio.transforms.Resample(orig_freq=sr,
new_freq=target_sr)
wav = transform(wav)
sr = target_sr
assert sr == target_sr
return wav.squeeze(0)
def save_audio(path: str,
tensor: torch.Tensor,
sr: int = 16000):
torchaudio.save(path, tensor.unsqueeze(0), sr)
def init_jit_model(model_path: str,
device=torch.device('cpu')):
torch.set_grad_enabled(False)
model = torch.jit.load(model_path, map_location=device)
model.eval()
return model
def get_speech_ts(wav: torch.Tensor,
model,
trig_sum: float = 0.25,
neg_trig_sum: float = 0.07,
num_steps: int = 8,
batch_size: int = 200,
num_samples_per_window: int = 4000,
min_speech_samples: int = 10000, #samples
min_silence_samples: int = 500,
run_function=validate,
visualize_probs=False):
num_samples = num_samples_per_window
assert num_samples % num_steps == 0
step = int(num_samples / num_steps) # stride / hop
outs = []
to_concat = []
for i in range(0, len(wav), step):
chunk = wav[i: i+num_samples]
if len(chunk) < num_samples:
chunk = F.pad(chunk, (0, num_samples - len(chunk)))
to_concat.append(chunk.unsqueeze(0))
if len(to_concat) >= batch_size:
chunks = torch.Tensor(torch.cat(to_concat, dim=0))
out = run_function(model, chunks)
outs.append(out)
to_concat = []
if to_concat:
chunks = torch.Tensor(torch.cat(to_concat, dim=0))
out = run_function(model, chunks)
outs.append(out)
outs = torch.cat(outs, dim=0)
buffer = deque(maxlen=num_steps) # maxlen reached => first element dropped
triggered = False
speeches = []
current_speech = {}
if visualize_probs:
import pandas as pd
smoothed_probs = []
speech_probs = outs[:, 1] # this is very misleading
temp_end = 0
for i, predict in enumerate(speech_probs): # add name
buffer.append(predict)
smoothed_prob = (sum(buffer) / len(buffer))
if visualize_probs:
smoothed_probs.append(float(smoothed_prob))
if (smoothed_prob >= trig_sum) and temp_end:
temp_end=0
if (smoothed_prob >= trig_sum) and not triggered:
triggered = True
current_speech['start'] = step * max(0, i-num_steps)
continue
if (smoothed_prob < neg_trig_sum) and triggered:
if not temp_end:
temp_end = step * i
if step * i - temp_end < min_silence_samples:
continue
else:
current_speech['end'] = temp_end
if (current_speech['end'] - current_speech['start']) > min_speech_samples:
speeches.append(current_speech)
temp_end = 0
current_speech = {}
triggered = False
continue
if current_speech:
current_speech['end'] = len(wav)
speeches.append(current_speech)
if visualize_probs:
pd.DataFrame({'probs':smoothed_probs}).plot(figsize=(16,8))
return speeches
def get_speech_ts_adaptive(wav: torch.Tensor,
model,
batch_size: int = 200,
step: int = 500,
num_samples_per_window: int = 4000, # Number of samples per audio chunk to feed to NN (4000 for 16k SR, 2000 for 8k SR is optimal)
min_speech_samples: int = 10000, # samples
min_silence_samples: int = 4000,
speech_pad_samples: int = 2000,
run_function=validate,
visualize_probs=False,
device='cpu'):
"""
This function is used for splitting long audios into speech chunks using silero VAD
Attention! All default sample rate values are optimal for 16000 sample rate model, if you are using 8000 sample rate model optimal values are half as much!
Parameters
----------
batch_size: int
batch size to feed to silero VAD (default - 200)
step: int
step size in samples, (default - 500)
num_samples_per_window: int
window size in samples (chunk length in samples to feed to NN, default - 4000)
min_speech_samples: int
if speech duration is shorter than this value, do not consider it speech (default - 10000)
min_silence_samples: int
number of samples to wait before considering as the end of speech (default - 4000)
speech_pad_samples: int
widen speech by this amount of samples each side (default - 2000)
run_function: function
function to use for the model call
visualize_probs: bool
whether draw prob hist or not (default: False)
device: string
torch device to use for the model call (default - "cpu")
Returns
----------
speeches: list
list containing ends and beginnings of speech chunks (in samples)
"""
num_samples = num_samples_per_window
num_steps = int(num_samples / step)
assert min_silence_samples >= step
outs = []
to_concat = []
for i in range(0, len(wav), step):
chunk = wav[i: i+num_samples]
if len(chunk) < num_samples:
chunk = F.pad(chunk, (0, num_samples - len(chunk)))
to_concat.append(chunk.unsqueeze(0))
if len(to_concat) >= batch_size:
chunks = torch.Tensor(torch.cat(to_concat, dim=0)).to(device)
out = run_function(model, chunks)
outs.append(out)
to_concat = []
if to_concat:
chunks = torch.Tensor(torch.cat(to_concat, dim=0)).to(device)
out = run_function(model, chunks)
outs.append(out)
outs = torch.cat(outs, dim=0).cpu()
buffer = deque(maxlen=num_steps)
triggered = False
speeches = []
smoothed_probs = []
current_speech = {}
speech_probs = outs[:, 1] # 0 index for silence probs, 1 index for speech probs
median_probs = speech_probs.median()
trig_sum = 0.89 * median_probs + 0.08 # 0.08 when median is zero, 0.97 when median is 1
temp_end = 0
for i, predict in enumerate(speech_probs):
buffer.append(predict)
smoothed_prob = max(buffer)
if visualize_probs:
smoothed_probs.append(float(smoothed_prob))
if (smoothed_prob >= trig_sum) and temp_end:
temp_end = 0
if (smoothed_prob >= trig_sum) and not triggered:
triggered = True
current_speech['start'] = step * max(0, i-num_steps)
continue
if (smoothed_prob < trig_sum) and triggered:
if not temp_end:
temp_end = step * i
if step * i - temp_end < min_silence_samples:
continue
else:
current_speech['end'] = temp_end
if (current_speech['end'] - current_speech['start']) > min_speech_samples:
speeches.append(current_speech)
temp_end = 0
current_speech = {}
triggered = False
continue
if current_speech:
current_speech['end'] = len(wav)
speeches.append(current_speech)
if visualize_probs:
pd.DataFrame({'probs': smoothed_probs}).plot(figsize=(16, 8))
for i, ts in enumerate(speeches):
if i == 0:
ts['start'] = max(0, ts['start'] - speech_pad_samples)
if i != len(speeches) - 1:
silence_duration = speeches[i+1]['start'] - ts['end']
if silence_duration < 2 * speech_pad_samples:
ts['end'] += silence_duration // 2
speeches[i+1]['start'] = max(0, speeches[i+1]['start'] - silence_duration // 2)
else:
ts['end'] += speech_pad_samples
else:
ts['end'] = min(len(wav), ts['end'] + speech_pad_samples)
return speeches
def get_number_ts(wav: torch.Tensor,
model,
model_stride=8,
hop_length=160,
sample_rate=16000,
run_function=validate):
wav = torch.unsqueeze(wav, dim=0)
perframe_logits = run_function(model, wav)[0]
perframe_preds = torch.argmax(torch.softmax(perframe_logits, dim=1), dim=1).squeeze() # (1, num_frames_strided)
extended_preds = []
for i in perframe_preds:
extended_preds.extend([i.item()] * model_stride)
# len(extended_preds) is *num_frames_real*; for each frame of audio we know if it has a number in it.
triggered = False
timings = []
cur_timing = {}
for i, pred in enumerate(extended_preds):
if pred == 1:
if not triggered:
cur_timing['start'] = int((i * hop_length) / (sample_rate / 1000))
triggered = True
elif pred == 0:
if triggered:
cur_timing['end'] = int((i * hop_length) / (sample_rate / 1000))
timings.append(cur_timing)
cur_timing = {}
triggered = False
if cur_timing:
cur_timing['end'] = int(len(wav) / (sample_rate / 1000))
timings.append(cur_timing)
return timings
def get_language(wav: torch.Tensor,
model,
run_function=validate):
wav = torch.unsqueeze(wav, dim=0)
lang_logits = run_function(model, wav)[2]
lang_pred = torch.argmax(torch.softmax(lang_logits, dim=1), dim=1).item() # from 0 to len(languages) - 1
assert lang_pred < len(languages)
return languages[lang_pred]
class VADiterator:
def __init__(self,
trig_sum: float = 0.26,
neg_trig_sum: float = 0.07,
num_steps: int = 8,
num_samples_per_window: int = 4000):
self.num_samples = num_samples_per_window
self.num_steps = num_steps
assert self.num_samples % num_steps == 0
self.step = int(self.num_samples / num_steps) # 500 samples is good enough
self.prev = torch.zeros(self.num_samples)
self.last = False
self.triggered = False
self.buffer = deque(maxlen=num_steps)
self.num_frames = 0
self.trig_sum = trig_sum
self.neg_trig_sum = neg_trig_sum
self.current_name = ''
def refresh(self):
self.prev = torch.zeros(self.num_samples)
self.last = False
self.triggered = False
self.buffer = deque(maxlen=self.num_steps)
self.num_frames = 0
def prepare_batch(self, wav_chunk, name=None):
if (name is not None) and (name != self.current_name):
self.refresh()
self.current_name = name
assert len(wav_chunk) <= self.num_samples
self.num_frames += len(wav_chunk)
if len(wav_chunk) < self.num_samples:
wav_chunk = F.pad(wav_chunk, (0, self.num_samples - len(wav_chunk))) # short chunk => eof audio
self.last = True
stacked = torch.cat([self.prev, wav_chunk])
self.prev = wav_chunk
overlap_chunks = [stacked[i:i+self.num_samples].unsqueeze(0)
for i in range(self.step, self.num_samples+1, self.step)]
return torch.cat(overlap_chunks, dim=0)
def state(self, model_out):
current_speech = {}
speech_probs = model_out[:, 1] # this is very misleading
for i, predict in enumerate(speech_probs):
self.buffer.append(predict)
if ((sum(self.buffer) / len(self.buffer)) >= self.trig_sum) and not self.triggered:
self.triggered = True
current_speech[self.num_frames - (self.num_steps-i) * self.step] = 'start'
if ((sum(self.buffer) / len(self.buffer)) < self.neg_trig_sum) and self.triggered:
current_speech[self.num_frames - (self.num_steps-i) * self.step] = 'end'
self.triggered = False
if self.triggered and self.last:
current_speech[self.num_frames] = 'end'
if self.last:
self.refresh()
return current_speech, self.current_name
class VADiteratorAdaptive:
def __init__(self,
trig_sum: float = 0.26,
neg_trig_sum: float = 0.06,
step: int = 500,
num_samples_per_window: int = 4000,
speech_pad_samples: int = 1000,
accum_period: int = 50):
"""
This class is used for streaming silero VAD usage
Parameters
----------
trig_sum: float
trigger value for speech probability, probs above this value are considered speech, switch to TRIGGERED state (default - 0.26)
neg_trig_sum: float
in triggered state probabilites below this value are considered nonspeech, switch to NONTRIGGERED state (default - 0.06)
step: int
step size in samples, (default - 500)
num_samples_per_window: int
window size in samples (chunk length in samples to feed to NN, default - 4000)
speech_pad_samples: int
widen speech by this amount of samples each side (default - 1000)
accum_period: int
number of chunks / iterations to wait before switching from constant (initial) trig and neg_trig coeffs to adaptive median coeffs (default - 50)
"""
self.num_samples = num_samples_per_window
self.num_steps = int(num_samples_per_window / step)
self.step = step
self.prev = torch.zeros(self.num_samples)
self.last = False
self.triggered = False
self.buffer = deque(maxlen=self.num_steps)
self.num_frames = 0
self.trig_sum = trig_sum
self.neg_trig_sum = neg_trig_sum
self.current_name = ''
self.median_meter = IterativeMedianMeter()
self.median = 0
self.total_steps = 0
self.accum_period = accum_period
self.speech_pad_samples = speech_pad_samples
def refresh(self):
self.prev = torch.zeros(self.num_samples)
self.last = False
self.triggered = False
self.buffer = deque(maxlen=self.num_steps)
self.num_frames = 0
self.median_meter.reset()
self.median = 0
self.total_steps = 0
def prepare_batch(self, wav_chunk, name=None):
if (name is not None) and (name != self.current_name):
self.refresh()
self.current_name = name
assert len(wav_chunk) <= self.num_samples
self.num_frames += len(wav_chunk)
if len(wav_chunk) < self.num_samples:
wav_chunk = F.pad(wav_chunk, (0, self.num_samples - len(wav_chunk))) # short chunk => eof audio
self.last = True
stacked = torch.cat([self.prev, wav_chunk])
self.prev = wav_chunk
overlap_chunks = [stacked[i:i+self.num_samples].unsqueeze(0)
for i in range(self.step, self.num_samples+1, self.step)]
return torch.cat(overlap_chunks, dim=0)
def state(self, model_out):
current_speech = {}
speech_probs = model_out[:, 1] # 0 index for silence probs, 1 index for speech probs
for i, predict in enumerate(speech_probs):
self.median = self.median_meter(predict.item())
if self.total_steps < self.accum_period:
trig_sum = self.trig_sum
neg_trig_sum = self.neg_trig_sum
else:
trig_sum = 0.89 * self.median + 0.08 # 0.08 when median is zero, 0.97 when median is 1
neg_trig_sum = 0.6 * self.median
self.total_steps += 1
self.buffer.append(predict)
smoothed_prob = max(self.buffer)
if (smoothed_prob >= trig_sum) and not self.triggered:
self.triggered = True
current_speech[max(0, self.num_frames - (self.num_steps-i) * self.step - self.speech_pad_samples)] = 'start'
if (smoothed_prob < neg_trig_sum) and self.triggered:
current_speech[self.num_frames - (self.num_steps-i) * self.step + self.speech_pad_samples] = 'end'
self.triggered = False
if self.triggered and self.last:
current_speech[self.num_frames] = 'end'
if self.last:
self.refresh()
return current_speech, self.current_name
def state_generator(model,
audios: List[str],
onnx: bool = False,
trig_sum: float = 0.26,
neg_trig_sum: float = 0.07,
num_steps: int = 8,
num_samples_per_window: int = 4000,
audios_in_stream: int = 2,
run_function=validate):
VADiters = [VADiterator(trig_sum, neg_trig_sum, num_steps, num_samples_per_window) for i in range(audios_in_stream)]
for i, current_pieces in enumerate(stream_imitator(audios, audios_in_stream, num_samples_per_window)):
for_batch = [x.prepare_batch(*y) for x, y in zip(VADiters, current_pieces)]
batch = torch.cat(for_batch)
outs = run_function(model, batch)
vad_outs = torch.split(outs, num_steps)
states = []
for x, y in zip(VADiters, vad_outs):
cur_st = x.state(y)
if cur_st[0]:
states.append(cur_st)
yield states
def stream_imitator(audios: List[str],
audios_in_stream: int,
num_samples_per_window: int = 4000):
audio_iter = iter(audios)
iterators = []
num_samples = num_samples_per_window
# initial wavs
for i in range(audios_in_stream):
next_wav = next(audio_iter)
wav = read_audio(next_wav)
wav_chunks = iter([(wav[i:i+num_samples], next_wav) for i in range(0, len(wav), num_samples)])
iterators.append(wav_chunks)
print('Done initial Loading')
good_iters = audios_in_stream
while True:
values = []
for i, it in enumerate(iterators):
try:
out, wav_name = next(it)
except StopIteration:
try:
next_wav = next(audio_iter)
print('Loading next wav: ', next_wav)
wav = read_audio(next_wav)
iterators[i] = iter([(wav[i:i+num_samples], next_wav) for i in range(0, len(wav), num_samples)])
out, wav_name = next(iterators[i])
except StopIteration:
good_iters -= 1
iterators[i] = repeat((torch.zeros(num_samples), 'junk'))
out, wav_name = next(iterators[i])
if good_iters == 0:
return
values.append((out, wav_name))
yield values
def single_audio_stream(model,
audio: torch.Tensor,
num_samples_per_window:int = 4000,
run_function=validate,
iterator_type='basic',
**kwargs):
num_samples = num_samples_per_window
if iterator_type == 'basic':
VADiter = VADiterator(num_samples_per_window=num_samples_per_window, **kwargs)
elif iterator_type == 'adaptive':
VADiter = VADiteratorAdaptive(num_samples_per_window=num_samples_per_window, **kwargs)
wav = read_audio(audio)
wav_chunks = iter([wav[i:i+num_samples] for i in range(0, len(wav), num_samples)])
for chunk in wav_chunks:
batch = VADiter.prepare_batch(chunk)
outs = run_function(model, batch)
states = []
state = VADiter.state(outs)
if state[0]:
states.append(state[0])
yield states
def collect_chunks(tss: List[dict],
wav: torch.Tensor):
chunks = []
for i in tss:
chunks.append(wav[i['start']: i['end']])
return torch.cat(chunks)
def drop_chunks(tss: List[dict],
wav: torch.Tensor):
chunks = []
cur_start = 0
for i in tss:
chunks.append((wav[cur_start: i['start']]))
cur_start = i['end']
return torch.cat(chunks)