-
Notifications
You must be signed in to change notification settings - Fork 229
/
vocoder.py
860 lines (745 loc) · 35.2 KB
/
vocoder.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
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
import os
import numpy as np
import yaml
import torch
import torch.nn.functional as F
import pyworld as pw
import parselmouth
import torchcrepe
import resampy
from transformers import HubertModel, Wav2Vec2FeatureExtractor
from fairseq import checkpoint_utils
from encoder.hubert.model import HubertSoft
from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
from torchaudio.transforms import Resample
from .unit2control import Unit2Control
from .core import frequency_filter, upsample, remove_above_fmax, MaskedAvgPool1d, MedianPool1d
import time
CREPE_RESAMPLE_KERNEL = {}
F0_KERNEL = {}
class F0_Extractor:
def __init__(self, f0_extractor, sample_rate = 44100, hop_size = 512, f0_min = 65, f0_max = 800):
self.f0_extractor = f0_extractor
self.sample_rate = sample_rate
self.hop_size = hop_size
self.f0_min = f0_min
self.f0_max = f0_max
if f0_extractor == 'crepe':
key_str = str(sample_rate)
if key_str not in CREPE_RESAMPLE_KERNEL:
CREPE_RESAMPLE_KERNEL[key_str] = Resample(sample_rate, 16000, lowpass_filter_width = 128)
self.resample_kernel = CREPE_RESAMPLE_KERNEL[key_str]
if f0_extractor == 'rmvpe':
if 'rmvpe' not in F0_KERNEL :
from encoder.rmvpe import RMVPE
F0_KERNEL['rmvpe'] = RMVPE('pretrain/rmvpe/model.pt', hop_length=160)
self.rmvpe = F0_KERNEL['rmvpe']
if f0_extractor == 'fcpe':
self.device_fcpe = 'cuda' if torch.cuda.is_available() else 'cpu'
if 'fcpe' not in F0_KERNEL :
from torchfcpe import spawn_bundled_infer_model
F0_KERNEL['fcpe'] = spawn_bundled_infer_model(device=self.device_fcpe)
self.fcpe = F0_KERNEL['fcpe']
def extract(self, audio, uv_interp = False, device = None, silence_front = 0): # audio: 1d numpy array
# extractor start time
n_frames = int(len(audio) // self.hop_size) + 1
start_frame = int(silence_front * self.sample_rate / self.hop_size)
real_silence_front = start_frame * self.hop_size / self.sample_rate
audio = audio[int(np.round(real_silence_front * self.sample_rate)) : ]
# extract f0 using parselmouth
if self.f0_extractor == 'parselmouth':
l_pad = int(np.ceil(1.5 / self.f0_min * self.sample_rate))
r_pad = int(self.hop_size * ((len(audio) - 1) // self.hop_size + 1) - len(audio) + l_pad + 1)
s = parselmouth.Sound(np.pad(audio, (l_pad, r_pad)), self.sample_rate).to_pitch_ac(
time_step = self.hop_size / self.sample_rate,
voicing_threshold = 0.6,
pitch_floor = self.f0_min,
pitch_ceiling = self.f0_max)
assert np.abs(s.t1 - 1.5 / self.f0_min) < 0.001
f0 = np.pad(s.selected_array['frequency'], (start_frame, 0))
if len(f0) < n_frames:
f0 = np.pad(f0, (0, n_frames - len(f0)))
f0 = f0[: n_frames]
# extract f0 using dio
elif self.f0_extractor == 'dio':
_f0, t = pw.dio(
audio.astype('double'),
self.sample_rate,
f0_floor = self.f0_min,
f0_ceil = self.f0_max,
channels_in_octave=2,
frame_period = (1000 * self.hop_size / self.sample_rate))
f0 = pw.stonemask(audio.astype('double'), _f0, t, self.sample_rate)
f0 = np.pad(f0.astype('float'), (start_frame, n_frames - len(f0) - start_frame))
# extract f0 using harvest
elif self.f0_extractor == 'harvest':
f0, _ = pw.harvest(
audio.astype('double'),
self.sample_rate,
f0_floor = self.f0_min,
f0_ceil = self.f0_max,
frame_period = (1000 * self.hop_size / self.sample_rate))
f0 = np.pad(f0.astype('float'), (start_frame, n_frames - len(f0) - start_frame))
# extract f0 using crepe
elif self.f0_extractor == 'crepe':
if device is None:
device = 'cuda' if torch.cuda.is_available() else 'cpu'
resample_kernel = self.resample_kernel.to(device)
wav16k_torch = resample_kernel(torch.FloatTensor(audio).unsqueeze(0).to(device))
f0, pd = torchcrepe.predict(wav16k_torch, 16000, 80, self.f0_min, self.f0_max, pad=True, model='full', batch_size=512, device=device, return_periodicity=True)
pd = MedianPool1d(pd, 4)
f0 = torchcrepe.threshold.At(0.05)(f0, pd)
f0 = MaskedAvgPool1d(f0, 4)
f0 = f0.squeeze(0).cpu().numpy()
f0 = np.array([f0[int(min(int(np.round(n * self.hop_size / self.sample_rate / 0.005)), len(f0) - 1))] for n in range(n_frames - start_frame)])
f0 = np.pad(f0, (start_frame, 0))
# extract f0 using rmvpe
elif self.f0_extractor == "rmvpe":
f0 = self.rmvpe.infer_from_audio(audio, self.sample_rate, device=device, thred=0.03, use_viterbi=False)
uv = f0 == 0
if len(f0[~uv]) > 0:
f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
origin_time = 0.01 * np.arange(len(f0))
target_time = self.hop_size / self.sample_rate * np.arange(n_frames - start_frame)
f0 = np.interp(target_time, origin_time, f0)
uv = np.interp(target_time, origin_time, uv.astype(float)) > 0.5
f0[uv] = 0
f0 = np.pad(f0, (start_frame, 0))
# extract f0 using fcpe
elif self.f0_extractor == "fcpe":
_audio = torch.from_numpy(audio).to(self.device_fcpe).unsqueeze(0)
f0 = self.fcpe(_audio, sr=self.sample_rate, decoder_mode="local_argmax", threshold=0.006)
f0 = f0.squeeze().cpu().numpy()
uv = f0 == 0
if len(f0[~uv]) > 0:
f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
origin_time = 0.01 * np.arange(len(f0))
target_time = self.hop_size / self.sample_rate * np.arange(n_frames - start_frame)
f0 = np.interp(target_time, origin_time, f0)
uv = np.interp(target_time, origin_time, uv.astype(float)) > 0.5
f0[uv] = 0
f0 = np.pad(f0, (start_frame, 0))
else:
raise ValueError(f" [x] Unknown f0 extractor: {self.f0_extractor}")
# interpolate the unvoiced f0
if uv_interp:
uv = f0 == 0
if len(f0[~uv]) > 0:
f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
f0[f0 < self.f0_min] = self.f0_min
return f0
class Volume_Extractor:
def __init__(self, hop_size = 512):
self.hop_size = hop_size
def extract(self, audio): # audio: 1d numpy array
n_frames = int(len(audio) // self.hop_size) + 1
audio2 = audio ** 2
audio2 = np.pad(audio2, (int(self.hop_size // 2), int((self.hop_size + 1) // 2)), mode = 'reflect')
volume = np.array([np.mean(audio2[int(n * self.hop_size) : int((n + 1) * self.hop_size)]) for n in range(n_frames)])
volume = np.sqrt(volume)
return volume
class Units_Encoder:
def __init__(self, encoder, encoder_ckpt, encoder_sample_rate = 16000, encoder_hop_size = 320, device = None,
cnhubertsoft_gate=10):
if device is None:
device = 'cuda' if torch.cuda.is_available() else 'cpu'
self.device = device
is_loaded_encoder = False
if encoder == 'hubertsoft':
self.model = Audio2HubertSoft(encoder_ckpt).to(device)
is_loaded_encoder = True
if encoder == 'hubertbase':
self.model = Audio2HubertBase(encoder_ckpt, device=device)
is_loaded_encoder = True
if encoder == 'hubertbase768':
self.model = Audio2HubertBase768(encoder_ckpt, device=device)
is_loaded_encoder = True
if encoder == 'hubertbase768l12':
self.model = Audio2HubertBase768L12(encoder_ckpt, device=device)
is_loaded_encoder = True
if encoder == 'hubertlarge1024l24':
self.model = Audio2HubertLarge1024L24(encoder_ckpt, device=device)
is_loaded_encoder = True
if encoder == 'contentvec':
self.model = Audio2ContentVec(encoder_ckpt, device=device)
is_loaded_encoder = True
if encoder == 'contentvec768':
self.model = Audio2ContentVec768(encoder_ckpt, device=device)
is_loaded_encoder = True
if encoder == 'contentvec768l12':
self.model = Audio2ContentVec768L12(encoder_ckpt, device=device)
is_loaded_encoder = True
if encoder == 'cnhubertsoftfish':
self.model = CNHubertSoftFish(encoder_ckpt, device=device, gate_size=cnhubertsoft_gate)
is_loaded_encoder = True
if not is_loaded_encoder:
raise ValueError(f" [x] Unknown units encoder: {encoder}")
self.resample_kernel = {}
self.encoder_sample_rate = encoder_sample_rate
self.encoder_hop_size = encoder_hop_size
def encode(self,
audio, # B, T
sample_rate,
hop_size):
# resample
if sample_rate == self.encoder_sample_rate:
audio_res = audio
else:
key_str = str(sample_rate)
if key_str not in self.resample_kernel:
self.resample_kernel[key_str] = Resample(sample_rate, self.encoder_sample_rate, lowpass_filter_width = 128).to(self.device)
audio_res = self.resample_kernel[key_str](audio)
# encode
if audio_res.size(-1) < 400:
audio_res = torch.nn.functional.pad(audio, (0, 400 - audio_res.size(-1)))
units = self.model(audio_res)
# alignment
n_frames = audio.size(-1) // hop_size + 1
ratio = (hop_size / sample_rate) / (self.encoder_hop_size / self.encoder_sample_rate)
index = torch.clamp(torch.round(ratio * torch.arange(n_frames).to(self.device)).long(), max = units.size(1) - 1)
units_aligned = torch.gather(units, 1, index.unsqueeze(0).unsqueeze(-1).repeat([1, 1, units.size(-1)]))
return units_aligned
class Audio2HubertSoft(torch.nn.Module):
def __init__(self, path, h_sample_rate = 16000, h_hop_size = 320):
super().__init__()
print(' [Encoder Model] HuBERT Soft')
self.hubert = HubertSoft()
print(' [Loading] ' + path)
checkpoint = torch.load(path)
consume_prefix_in_state_dict_if_present(checkpoint, "module.")
self.hubert.load_state_dict(checkpoint)
self.hubert.eval()
def forward(self,
audio): # B, T
with torch.inference_mode():
units = self.hubert.units(audio.unsqueeze(1))
return units
class Audio2ContentVec():
def __init__(self, path, h_sample_rate=16000, h_hop_size=320, device='cpu'):
self.device = device
print(' [Encoder Model] Content Vec')
print(' [Loading] ' + path)
self.models, self.saved_cfg, self.task = checkpoint_utils.load_model_ensemble_and_task([path], suffix="", )
self.hubert = self.models[0]
self.hubert = self.hubert.to(self.device)
self.hubert.eval()
def __call__(self,
audio): # B, T
# wav_tensor = torch.from_numpy(audio).to(self.device)
wav_tensor = audio
feats = wav_tensor.view(1, -1)
padding_mask = torch.BoolTensor(feats.shape).fill_(False)
inputs = {
"source": feats.to(wav_tensor.device),
"padding_mask": padding_mask.to(wav_tensor.device),
"output_layer": 9, # layer 9
}
with torch.no_grad():
logits = self.hubert.extract_features(**inputs)
feats = self.hubert.final_proj(logits[0])
units = feats # .transpose(2, 1)
return units
class Audio2ContentVec768():
def __init__(self, path, h_sample_rate=16000, h_hop_size=320, device='cpu'):
self.device = device
print(' [Encoder Model] Content Vec')
print(' [Loading] ' + path)
self.models, self.saved_cfg, self.task = checkpoint_utils.load_model_ensemble_and_task([path], suffix="", )
self.hubert = self.models[0]
self.hubert = self.hubert.to(self.device)
self.hubert.eval()
def __call__(self,
audio): # B, T
# wav_tensor = torch.from_numpy(audio).to(self.device)
wav_tensor = audio
feats = wav_tensor.view(1, -1)
padding_mask = torch.BoolTensor(feats.shape).fill_(False)
inputs = {
"source": feats.to(wav_tensor.device),
"padding_mask": padding_mask.to(wav_tensor.device),
"output_layer": 9, # layer 9
}
with torch.no_grad():
logits = self.hubert.extract_features(**inputs)
feats = logits[0]
units = feats # .transpose(2, 1)
return units
class Audio2ContentVec768L12():
def __init__(self, path, h_sample_rate=16000, h_hop_size=320, device='cpu'):
self.device = device
print(' [Encoder Model] Content Vec')
print(' [Loading] ' + path)
self.models, self.saved_cfg, self.task = checkpoint_utils.load_model_ensemble_and_task([path], suffix="", )
self.hubert = self.models[0]
self.hubert = self.hubert.to(self.device)
self.hubert.eval()
def __call__(self,
audio): # B, T
# wav_tensor = torch.from_numpy(audio).to(self.device)
wav_tensor = audio
feats = wav_tensor.view(1, -1)
padding_mask = torch.BoolTensor(feats.shape).fill_(False)
inputs = {
"source": feats.to(wav_tensor.device),
"padding_mask": padding_mask.to(wav_tensor.device),
"output_layer": 12, # layer 12
}
with torch.no_grad():
logits = self.hubert.extract_features(**inputs)
feats = logits[0]
units = feats # .transpose(2, 1)
return units
class CNHubertSoftFish(torch.nn.Module):
def __init__(self, path, h_sample_rate=16000, h_hop_size=320, device='cpu', gate_size=10):
super().__init__()
self.device = device
self.gate_size = gate_size
self.feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(
"./pretrain/TencentGameMate/chinese-hubert-base")
self.model = HubertModel.from_pretrained("./pretrain/TencentGameMate/chinese-hubert-base")
self.proj = torch.nn.Sequential(torch.nn.Dropout(0.1), torch.nn.Linear(768, 256))
# self.label_embedding = nn.Embedding(128, 256)
state_dict = torch.load(path, map_location=device)
self.load_state_dict(state_dict)
@torch.no_grad()
def forward(self, audio):
input_values = self.feature_extractor(
audio, sampling_rate=16000, return_tensors="pt"
).input_values
input_values = input_values.to(self.model.device)
return self._forward(input_values[0])
@torch.no_grad()
def _forward(self, input_values):
features = self.model(input_values)
features = self.proj(features.last_hidden_state)
# Top-k gating
topk, indices = torch.topk(features, self.gate_size, dim=2)
features = torch.zeros_like(features).scatter(2, indices, topk)
features = features / features.sum(2, keepdim=True)
return features.to(self.device) # .transpose(1, 2)
class Audio2HubertBase():
def __init__(self, path, h_sample_rate=16000, h_hop_size=320, device='cpu'):
self.device = device
print(' [Encoder Model] HuBERT Base')
print(' [Loading] ' + path)
self.models, self.saved_cfg, self.task = checkpoint_utils.load_model_ensemble_and_task([path], suffix="", )
self.hubert = self.models[0]
self.hubert = self.hubert.to(self.device)
self.hubert = self.hubert.float()
self.hubert.eval()
def __call__(self,
audio): # B, T
with torch.no_grad():
padding_mask = torch.BoolTensor(audio.shape).fill_(False)
inputs = {
"source": audio.to(self.device),
"padding_mask": padding_mask.to(self.device),
"output_layer": 9, # layer 9
}
logits = self.hubert.extract_features(**inputs)
units = self.hubert.final_proj(logits[0])
return units
class Audio2HubertBase768():
def __init__(self, path, h_sample_rate=16000, h_hop_size=320, device='cpu'):
self.device = device
print(' [Encoder Model] HuBERT Base')
print(' [Loading] ' + path)
self.models, self.saved_cfg, self.task = checkpoint_utils.load_model_ensemble_and_task([path], suffix="", )
self.hubert = self.models[0]
self.hubert = self.hubert.to(self.device)
self.hubert = self.hubert.float()
self.hubert.eval()
def __call__(self,
audio): # B, T
with torch.no_grad():
padding_mask = torch.BoolTensor(audio.shape).fill_(False)
inputs = {
"source": audio.to(self.device),
"padding_mask": padding_mask.to(self.device),
"output_layer": 9, # layer 9
}
logits = self.hubert.extract_features(**inputs)
units = logits[0]
return units
class Audio2HubertBase768L12():
def __init__(self, path, h_sample_rate=16000, h_hop_size=320, device='cpu'):
self.device = device
print(' [Encoder Model] HuBERT Base')
print(' [Loading] ' + path)
self.models, self.saved_cfg, self.task = checkpoint_utils.load_model_ensemble_and_task([path], suffix="", )
self.hubert = self.models[0]
self.hubert = self.hubert.to(self.device)
self.hubert = self.hubert.float()
self.hubert.eval()
def __call__(self,
audio): # B, T
with torch.no_grad():
padding_mask = torch.BoolTensor(audio.shape).fill_(False)
inputs = {
"source": audio.to(self.device),
"padding_mask": padding_mask.to(self.device),
"output_layer": 12, # layer 12
}
logits = self.hubert.extract_features(**inputs)
units = logits[0]
return units
class Audio2HubertLarge1024L24():
def __init__(self, path, h_sample_rate=16000, h_hop_size=320, device='cpu'):
self.device = device
print(' [Encoder Model] HuBERT Base')
print(' [Loading] ' + path)
self.models, self.saved_cfg, self.task = checkpoint_utils.load_model_ensemble_and_task([path], suffix="", )
self.hubert = self.models[0]
self.hubert = self.hubert.to(self.device)
self.hubert = self.hubert.float()
self.hubert.eval()
def __call__(self,
audio): # B, T
with torch.no_grad():
padding_mask = torch.BoolTensor(audio.shape).fill_(False)
inputs = {
"source": audio.to(self.device),
"padding_mask": padding_mask.to(self.device),
"output_layer": 24, # layer 24
}
logits = self.hubert.extract_features(**inputs)
units = logits[0]
return units
class DotDict(dict):
def __getattr__(*args):
val = dict.get(*args)
return DotDict(val) if type(val) is dict else val
__setattr__ = dict.__setitem__
__delattr__ = dict.__delitem__
def load_model(
model_path,
device='cpu'):
config_file = os.path.join(os.path.split(model_path)[0], 'config.yaml')
with open(config_file, "r") as config:
args = yaml.safe_load(config)
args = DotDict(args)
# load model
model = None
if args.model.type == 'Sins':
model = Sins(
sampling_rate=args.data.sampling_rate,
block_size=args.data.block_size,
n_harmonics=args.model.n_harmonics,
n_mag_allpass=args.model.n_mag_allpass,
n_mag_noise=args.model.n_mag_noise,
n_unit=args.data.encoder_out_channels,
n_spk=args.model.n_spk)
elif args.model.type == 'CombSub':
model = CombSub(
sampling_rate=args.data.sampling_rate,
block_size=args.data.block_size,
n_mag_allpass=args.model.n_mag_allpass,
n_mag_harmonic=args.model.n_mag_harmonic,
n_mag_noise=args.model.n_mag_noise,
n_unit=args.data.encoder_out_channels,
n_spk=args.model.n_spk)
elif args.model.type == 'CombSubFast':
model = CombSubFast(
sampling_rate=args.data.sampling_rate,
block_size=args.data.block_size,
n_unit=args.data.encoder_out_channels,
n_spk=args.model.n_spk)
elif args.model.type == 'CombSubSuperFast':
model = CombSubSuperFast(
sampling_rate=args.data.sampling_rate,
block_size=args.data.block_size,
win_length=args.model.win_length,
n_unit=args.data.encoder_out_channels,
n_spk=args.model.n_spk)
else:
raise ValueError(f" [x] Unknown Model: {args.model.type}")
print(' [Loading] ' + model_path)
ckpt = torch.load(model_path, map_location=torch.device(device))
model.to(device)
model.load_state_dict(ckpt['model'])
model.eval()
return model, args
class Sins(torch.nn.Module):
def __init__(self,
sampling_rate,
block_size,
n_harmonics,
n_mag_allpass,
n_mag_noise,
n_unit=256,
n_spk=1):
super().__init__()
print(' [DDSP Model] Sinusoids Additive Synthesiser')
# params
self.register_buffer("sampling_rate", torch.tensor(sampling_rate))
self.register_buffer("block_size", torch.tensor(block_size))
# Unit2Control
split_map = {
'amplitudes': n_harmonics,
'group_delay': n_mag_allpass,
'noise_magnitude': n_mag_noise,
}
self.unit2ctrl = Unit2Control(n_unit, n_spk, split_map)
def forward(self, units_frames, f0_frames, volume_frames, spk_id=None, spk_mix_dict=None, initial_phase=None, infer=True, max_upsample_dim=32):
'''
units_frames: B x n_frames x n_unit
f0_frames: B x n_frames x 1
volume_frames: B x n_frames x 1
spk_id: B x 1
'''
# exciter phase
f0 = upsample(f0_frames, self.block_size)
if infer:
x = torch.cumsum(f0.double() / self.sampling_rate, axis=1)
else:
x = torch.cumsum(f0 / self.sampling_rate, axis=1)
if initial_phase is not None:
x += initial_phase.to(x) / 2 / np.pi
x = x - torch.round(x)
x = x.to(f0)
phase = 2 * np.pi * x
phase_frames = phase[:, ::self.block_size, :]
# parameter prediction
ctrls, hidden = self.unit2ctrl(units_frames, f0_frames, phase_frames, volume_frames, spk_id=spk_id, spk_mix_dict=spk_mix_dict)
amplitudes_frames = torch.exp(ctrls['amplitudes'])/ 128
group_delay = np.pi * torch.tanh(ctrls['group_delay'])
noise_param = torch.exp(ctrls['noise_magnitude']) / 128
# sinusoids exciter signal
amplitudes_frames = remove_above_fmax(amplitudes_frames, f0_frames, self.sampling_rate / 2, level_start = 1)
n_harmonic = amplitudes_frames.shape[-1]
level_harmonic = torch.arange(1, n_harmonic + 1).to(phase)
sinusoids = 0.
for n in range(( n_harmonic - 1) // max_upsample_dim + 1):
start = n * max_upsample_dim
end = (n + 1) * max_upsample_dim
phases = phase * level_harmonic[start:end]
amplitudes = upsample(amplitudes_frames[:,:,start:end], self.block_size)
sinusoids += (torch.sin(phases) * amplitudes).sum(-1)
# harmonic part filter (apply group-delay)
harmonic = frequency_filter(
sinusoids,
torch.exp(1.j * torch.cumsum(group_delay, axis = -1)),
hann_window = False)
# noise part filter
noise = torch.rand_like(harmonic) * 2 - 1
noise = frequency_filter(
noise,
torch.complex(noise_param, torch.zeros_like(noise_param)),
hann_window = True)
signal = harmonic + noise
return signal, hidden, (harmonic, noise) #, (noise_param, noise_param)
class CombSubSuperFast(torch.nn.Module):
def __init__(self,
sampling_rate,
block_size,
win_length,
n_unit=256,
n_spk=1,
use_pitch_aug=False,
pcmer_norm=False):
super().__init__()
print(' [DDSP Model] Combtooth Subtractive Synthesiser')
# params
self.register_buffer("sampling_rate", torch.tensor(sampling_rate))
self.register_buffer("block_size", torch.tensor(block_size))
self.register_buffer("win_length", torch.tensor(win_length))
self.register_buffer("window", torch.hann_window(win_length))
#Unit2Control
split_map = {
'harmonic_magnitude': win_length // 2 + 1,
'harmonic_phase': win_length // 2 + 1,
'noise_magnitude': win_length // 2 + 1,
'noise_phase': win_length // 2 + 1
}
self.unit2ctrl = Unit2Control(n_unit, n_spk, split_map, use_pitch_aug=use_pitch_aug, use_naive_v2=True)
def forward(self, units_frames, f0_frames, volume_frames, spk_id=None, spk_mix_dict=None, aug_shift=None, initial_phase=None, infer=True, **kwargs):
'''
units_frames: B x n_frames x n_unit
f0_frames: B x n_frames x 1
volume_frames: B x n_frames x 1
spk_id: B x 1
'''
# exciter phase
f0 = upsample(f0_frames, self.block_size)
if infer:
x = torch.cumsum(f0.double() / self.sampling_rate, axis=1)
else:
x = torch.cumsum(f0 / self.sampling_rate, axis=1)
if initial_phase is not None:
x += initial_phase.to(x) / 2 / np.pi
x = x - torch.round(x)
x = x.to(f0)
phase_frames = 2 * np.pi * x[:, ::self.block_size, :]
# parameter prediction
ctrls, hidden = self.unit2ctrl(units_frames, f0_frames, phase_frames, volume_frames, spk_id=spk_id, spk_mix_dict=spk_mix_dict, aug_shift=aug_shift)
src_filter = torch.exp(ctrls['harmonic_magnitude'] + 1.j * np.pi * ctrls['harmonic_phase'])
src_filter = torch.cat((src_filter, src_filter[:,-1:,:]), 1)
noise_filter= torch.exp(ctrls['noise_magnitude'] + 1.j * np.pi * ctrls['noise_phase']) / 128
noise_filter = torch.cat((noise_filter, noise_filter[:,-1:,:]), 1)
# combtooth exciter signal
combtooth = torch.sinc(self.sampling_rate * x / (f0 + 1e-3))
combtooth = combtooth.squeeze(-1)
if combtooth.shape[-1] > self.win_length // 2:
pad_mode = 'reflect'
else:
pad_mode = 'constant'
combtooth_stft = torch.stft(
combtooth,
n_fft = self.win_length,
win_length = self.win_length,
hop_length = self.block_size,
window = self.window,
center = True,
return_complex = True,
pad_mode = pad_mode)
# noise exciter signal
noise = torch.randn_like(combtooth)
noise_stft = torch.stft(
noise,
n_fft = self.win_length,
win_length = self.win_length,
hop_length = self.block_size,
window = self.window,
center = True,
return_complex = True,
pad_mode = pad_mode)
# apply the filters
signal_stft = combtooth_stft * src_filter.permute(0, 2, 1) + noise_stft * noise_filter.permute(0, 2, 1)
# take the istft to resynthesize audio.
signal = torch.istft(
signal_stft,
n_fft = self.win_length,
win_length = self.win_length,
hop_length = self.block_size,
window = self.window,
center = True)
return signal, hidden, (signal, signal)
class CombSubFast(torch.nn.Module):
def __init__(self,
sampling_rate,
block_size,
n_unit=256,
n_spk=1,
use_pitch_aug=False,
pcmer_norm=False):
super().__init__()
print(' [DDSP Model] Combtooth Subtractive Synthesiser')
# params
self.register_buffer("sampling_rate", torch.tensor(sampling_rate))
self.register_buffer("block_size", torch.tensor(block_size))
self.register_buffer("window", torch.sqrt(torch.hann_window(2 * block_size)))
#Unit2Control
split_map = {
'harmonic_magnitude': block_size + 1,
'harmonic_phase': block_size + 1,
'noise_magnitude': block_size + 1
}
self.unit2ctrl = Unit2Control(n_unit, n_spk, split_map, use_pitch_aug=use_pitch_aug, pcmer_norm=pcmer_norm)
def forward(self, units_frames, f0_frames, volume_frames, spk_id=None, spk_mix_dict=None, aug_shift=None, initial_phase=None, infer=True, **kwargs):
'''
units_frames: B x n_frames x n_unit
f0_frames: B x n_frames x 1
volume_frames: B x n_frames x 1
spk_id: B x 1
'''
# exciter phase
f0 = upsample(f0_frames, self.block_size)
if infer:
x = torch.cumsum(f0.double() / self.sampling_rate, axis=1)
else:
x = torch.cumsum(f0 / self.sampling_rate, axis=1)
if initial_phase is not None:
x += initial_phase.to(x) / 2 / np.pi
x = x - torch.round(x)
x = x.to(f0)
phase_frames = 2 * np.pi * x[:, ::self.block_size, :]
# parameter prediction
ctrls, hidden = self.unit2ctrl(units_frames, f0_frames, phase_frames, volume_frames, spk_id=spk_id, spk_mix_dict=spk_mix_dict, aug_shift=aug_shift)
src_filter = torch.exp(ctrls['harmonic_magnitude'] + 1.j * np.pi * ctrls['harmonic_phase'])
src_filter = torch.cat((src_filter, src_filter[:,-1:,:]), 1)
noise_filter= torch.exp(ctrls['noise_magnitude']) / 128
noise_filter = torch.cat((noise_filter, noise_filter[:,-1:,:]), 1)
# combtooth exciter signal
combtooth = torch.sinc(self.sampling_rate * x / (f0 + 1e-3))
combtooth = combtooth.squeeze(-1)
combtooth_frames = F.pad(combtooth, (self.block_size, self.block_size)).unfold(1, 2 * self.block_size, self.block_size)
combtooth_frames = combtooth_frames * self.window
combtooth_fft = torch.fft.rfft(combtooth_frames, 2 * self.block_size)
# noise exciter signal
noise = torch.rand_like(combtooth) * 2 - 1
noise_frames = F.pad(noise, (self.block_size, self.block_size)).unfold(1, 2 * self.block_size, self.block_size)
noise_frames = noise_frames * self.window
noise_fft = torch.fft.rfft(noise_frames, 2 * self.block_size)
# apply the filters
signal_fft = combtooth_fft * src_filter + noise_fft * noise_filter
# take the ifft to resynthesize audio.
signal_frames_out = torch.fft.irfft(signal_fft, 2 * self.block_size) * self.window
# overlap add
fold = torch.nn.Fold(output_size=(1, (signal_frames_out.size(1) + 1) * self.block_size), kernel_size=(1, 2 * self.block_size), stride=(1, self.block_size))
signal = fold(signal_frames_out.transpose(1, 2))[:, 0, 0, self.block_size : -self.block_size]
return signal, hidden, (signal, signal)
class CombSub(torch.nn.Module):
def __init__(self,
sampling_rate,
block_size,
n_mag_allpass,
n_mag_harmonic,
n_mag_noise,
n_unit=256,
n_spk=1):
super().__init__()
print(' [DDSP Model] Combtooth Subtractive Synthesiser (Old Version)')
# params
self.register_buffer("sampling_rate", torch.tensor(sampling_rate))
self.register_buffer("block_size", torch.tensor(block_size))
#Unit2Control
split_map = {
'group_delay': n_mag_allpass,
'harmonic_magnitude': n_mag_harmonic,
'noise_magnitude': n_mag_noise
}
self.unit2ctrl = Unit2Control(n_unit, n_spk, split_map)
def forward(self, units_frames, f0_frames, volume_frames, spk_id=None, spk_mix_dict=None, initial_phase=None, infer=True, **kwargs):
'''
units_frames: B x n_frames x n_unit
f0_frames: B x n_frames x 1
volume_frames: B x n_frames x 1
spk_id: B x 1
'''
# exciter phase
f0 = upsample(f0_frames, self.block_size)
if infer:
x = torch.cumsum(f0.double() / self.sampling_rate, axis=1)
else:
x = torch.cumsum(f0 / self.sampling_rate, axis=1)
if initial_phase is not None:
x += initial_phase.to(x) / 2 / np.pi
x = x - torch.round(x)
x = x.to(f0)
phase_frames = 2 * np.pi * x[:, ::self.block_size, :]
# parameter prediction
ctrls, hidden = self.unit2ctrl(units_frames, f0_frames, phase_frames, volume_frames, spk_id=spk_id, spk_mix_dict=spk_mix_dict)
group_delay = np.pi * torch.tanh(ctrls['group_delay'])
src_param = torch.exp(ctrls['harmonic_magnitude'])
noise_param = torch.exp(ctrls['noise_magnitude']) / 128
# combtooth exciter signal
combtooth = torch.sinc(self.sampling_rate * x / (f0 + 1e-3))
combtooth = combtooth.squeeze(-1)
# harmonic part filter (using dynamic-windowed LTV-FIR, with group-delay prediction)
harmonic = frequency_filter(
combtooth,
torch.exp(1.j * torch.cumsum(group_delay, axis = -1)),
hann_window = False)
harmonic = frequency_filter(
harmonic,
torch.complex(src_param, torch.zeros_like(src_param)),
hann_window = True,
half_width_frames = 1.5 * self.sampling_rate / (f0_frames + 1e-3))
# noise part filter (using constant-windowed LTV-FIR, without group-delay)
noise = torch.rand_like(harmonic) * 2 - 1
noise = frequency_filter(
noise,
torch.complex(noise_param, torch.zeros_like(noise_param)),
hann_window = True)
signal = harmonic + noise
return signal, hidden, (harmonic, noise)