forked from speechbrain/speechbrain
-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathfeatures.py
More file actions
1916 lines (1628 loc) · 61 KB
/
Copy pathfeatures.py
File metadata and controls
1916 lines (1628 loc) · 61 KB
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
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
"""Low-level feature pipeline components
This library gathers functions that compute popular speech features over
batches of data. All the classes are of type nn.Module. This gives the
possibility to have end-to-end differentiability and to backpropagate the
gradient through them. Our functions are a modified version the ones
in torch audio toolkit (https://github.com/pytorch/audio).
Example
-------
>>> import torch
>>> from speechbrain.dataio.dataio import read_audio
>>> signal =read_audio('tests/samples/single-mic/example1.wav')
>>> signal = signal.unsqueeze(0)
>>> compute_STFT = STFT(
... sample_rate=16000, win_length=25, hop_length=10, n_fft=400
... )
>>> features = compute_STFT(signal)
>>> features = spectral_magnitude(features)
>>> compute_fbanks = Filterbank(n_mels=40)
>>> features = compute_fbanks(features)
>>> compute_mfccs = DCT(input_size=40, n_out=20)
>>> features = compute_mfccs(features)
>>> compute_deltas = Deltas(input_size=20)
>>> delta1 = compute_deltas(features)
>>> delta2 = compute_deltas(delta1)
>>> features = torch.cat([features, delta1, delta2], dim=2)
>>> compute_cw = ContextWindow(left_frames=5, right_frames=5)
>>> features = compute_cw(features)
>>> norm = InputNormalization()
>>> features = norm(features, torch.tensor([1]).float())
Authors
* Mirco Ravanelli 2020
* Peter Plantinga 2025
* Rogier van Dalen 2025
"""
import math
from typing import Optional, Tuple, Union
import torch
from torch.distributed import ReduceOp
from speechbrain.utils.checkpoints import (
mark_as_loader,
mark_as_saver,
mark_as_transfer,
register_checkpoint_hooks,
)
from speechbrain.utils.distributed import ddp_all_reduce
from speechbrain.utils.filter_analysis import FilterProperties
from speechbrain.utils.logger import get_logger
logger = get_logger(__name__)
class STFT(torch.nn.Module):
"""computes the Short-Term Fourier Transform (STFT).
This class computes the Short-Term Fourier Transform of an audio signal.
It supports multi-channel audio inputs (batch, time, channels).
Arguments
---------
sample_rate : int
Sample rate of the input audio signal (e.g 16000).
win_length : float
Length (in ms) of the sliding window used to compute the STFT.
hop_length : float
Length (in ms) of the hope of the sliding window used to compute
the STFT.
n_fft : int
Number of fft point of the STFT. It defines the frequency resolution
(n_fft should be <= than win_len).
window_fn : function
A function that takes an integer (number of samples) and outputs a
tensor to be multiplied with each window before fft.
normalized_stft : bool
If True, the function returns the normalized STFT results,
i.e., multiplied by win_length^-0.5 (default is False).
center : bool
If True (default), the input will be padded on both sides so that the
t-th frame is centered at time t×hop_length. Otherwise, the t-th frame
begins at time t×hop_length.
pad_mode : str
It can be 'constant','reflect','replicate', 'circular', 'reflect'
(default). 'constant' pads the input tensor boundaries with a
constant value. 'reflect' pads the input tensor using the reflection
of the input boundary. 'replicate' pads the input tensor using
replication of the input boundary. 'circular' pads using circular
replication.
onesided : True
If True (default) only returns nfft/2 values. Note that the other
samples are redundant due to the Fourier transform conjugate symmetry.
Example
-------
>>> import torch
>>> compute_STFT = STFT(
... sample_rate=16000, win_length=25, hop_length=10, n_fft=400
... )
>>> inputs = torch.randn([10, 16000])
>>> features = compute_STFT(inputs)
>>> features.shape
torch.Size([10, 101, 201, 2])
"""
def __init__(
self,
sample_rate,
win_length=25,
hop_length=10,
n_fft=400,
window_fn=torch.hamming_window,
normalized_stft=False,
center=True,
pad_mode="constant",
onesided=True,
):
super().__init__()
self.sample_rate = sample_rate
self.win_length = win_length
self.hop_length = hop_length
self.n_fft = n_fft
self.normalized_stft = normalized_stft
self.center = center
self.pad_mode = pad_mode
self.onesided = onesided
# Convert win_length and hop_length from ms to samples
self.win_length = int(
round((self.sample_rate / 1000.0) * self.win_length)
)
self.hop_length = int(
round((self.sample_rate / 1000.0) * self.hop_length)
)
self.window = window_fn(self.win_length)
def forward(self, x):
"""Returns the STFT generated from the input waveforms.
Arguments
---------
x : torch.Tensor
A batch of audio signals to transform.
Returns
-------
stft : torch.Tensor
"""
# Managing multi-channel stft
or_shape = x.shape
if len(or_shape) == 3:
x = x.transpose(1, 2)
x = x.reshape(or_shape[0] * or_shape[2], or_shape[1])
stft = torch.stft(
x,
self.n_fft,
self.hop_length,
self.win_length,
self.window.to(x.device),
self.center,
self.pad_mode,
self.normalized_stft,
self.onesided,
return_complex=True,
)
stft = torch.view_as_real(stft)
# Retrieving the original dimensionality (batch,time, channels)
if len(or_shape) == 3:
stft = stft.reshape(
or_shape[0],
or_shape[2],
stft.shape[1],
stft.shape[2],
stft.shape[3],
)
stft = stft.permute(0, 3, 2, 4, 1)
else:
# (batch, time, channels)
stft = stft.transpose(2, 1)
return stft
def get_filter_properties(self) -> FilterProperties:
if not self.center:
raise ValueError(
"ValueProperties cannot model a non-centered STFT, as it "
"assumes either centering or causality"
)
return FilterProperties(
window_size=self.win_length, stride=self.hop_length
)
class ISTFT(torch.nn.Module):
"""Computes the Inverse Short-Term Fourier Transform (ISTFT)
This class computes the Inverse Short-Term Fourier Transform of
an audio signal. It supports multi-channel audio inputs
(batch, time_step, n_fft, 2, n_channels [optional]).
Arguments
---------
sample_rate : int
Sample rate of the input audio signal (e.g. 16000).
n_fft : int
Number of points in FFT.
win_length : float
Length (in ms) of the sliding window used when computing the STFT.
hop_length : float
Length (in ms) of the hope of the sliding window used when computing
the STFT.
window_fn : function
A function that takes an integer (number of samples) and outputs a
tensor to be used as a window for ifft.
normalized_stft : bool
If True, the function assumes that it's working with the normalized
STFT results. (default is False)
center : bool
If True (default), the function assumes that the STFT result was padded
on both sides.
onesided : True
If True (default), the function assumes that there are n_fft/2 values
for each time frame of the STFT.
epsilon : float
A small value to avoid division by 0 when normalizing by the sum of the
squared window. Playing with it can fix some abnormalities at the
beginning and at the end of the reconstructed signal. The default value
of epsilon is 1e-12.
Example
-------
>>> import torch
>>> compute_STFT = STFT(
... sample_rate=16000, win_length=25, hop_length=10, n_fft=400
... )
>>> compute_ISTFT = ISTFT(
... sample_rate=16000, win_length=25, hop_length=10
... )
>>> inputs = torch.randn([10, 16000])
>>> outputs = compute_ISTFT(compute_STFT(inputs))
>>> outputs.shape
torch.Size([10, 16000])
"""
def __init__(
self,
sample_rate,
n_fft=None,
win_length=25,
hop_length=10,
window_fn=torch.hamming_window,
normalized_stft=False,
center=True,
onesided=True,
epsilon=1e-12,
):
super().__init__()
self.sample_rate = sample_rate
self.n_fft = n_fft
self.win_length = win_length
self.hop_length = hop_length
self.normalized_stft = normalized_stft
self.center = center
self.onesided = onesided
self.epsilon = epsilon
# Convert win_length and hop_length from ms to samples
self.win_length = int(
round((self.sample_rate / 1000.0) * self.win_length)
)
self.hop_length = int(
round((self.sample_rate / 1000.0) * self.hop_length)
)
# Create window using provided function
self.window = window_fn(self.win_length)
def forward(self, x, sig_length=None):
"""Returns the ISTFT generated from the input signal.
Arguments
---------
x : torch.Tensor
A batch of audio signals in the frequency domain to transform.
sig_length : int
The length of the output signal in number of samples. If not
specified will be equal to: (time_step - 1) * hop_length + n_fft
Returns
-------
istft : torch.Tensor
"""
or_shape = x.shape
# Infer n_fft if not provided
if self.n_fft is None and self.onesided:
n_fft = (x.shape[2] - 1) * 2
elif self.n_fft is None and not self.onesided:
n_fft = x.shape[2]
else:
n_fft = self.n_fft
# Changing the format for (batch, time_step, n_fft, 2, n_channels)
if len(or_shape) == 5:
x = x.permute(0, 4, 2, 1, 3)
# Lumping batch and channel dimension, because torch.istft
# doesn't support batching.
x = x.reshape(-1, x.shape[2], x.shape[3], x.shape[4])
elif len(or_shape) == 4:
x = x.permute(0, 2, 1, 3)
# isft ask complex input
x = torch.complex(x[..., 0], x[..., 1])
istft = torch.istft(
input=x,
n_fft=n_fft,
hop_length=self.hop_length,
win_length=self.win_length,
window=self.window.to(x.device),
center=self.center,
onesided=self.onesided,
length=sig_length,
)
# Convert back to (time, time_step, n_channels)
if len(or_shape) == 5:
istft = istft.reshape(or_shape[0], or_shape[4], -1)
istft = istft.transpose(1, 2)
return istft
def spectral_magnitude(
stft, power: float = 1, log: bool = False, eps: float = 1e-14
):
"""Returns the magnitude of a complex spectrogram.
Arguments
---------
stft : torch.Tensor
A tensor, output from the stft function.
power : int
What power to use in computing the magnitude.
Use power=1 for the power spectrogram.
Use power=0.5 for the magnitude spectrogram.
log : bool
Whether to apply log to the spectral features.
eps : float
A small value to prevent square root of zero.
Returns
-------
spectr : torch.Tensor
Example
-------
>>> a = torch.Tensor([[3, 4]])
>>> spectral_magnitude(a, power=0.5)
tensor([5.])
"""
spectr = stft.pow(2).sum(-1)
# Add eps avoids NaN when spectr is zero
if power < 1:
spectr = spectr + eps
spectr = spectr.pow(power)
if log:
return torch.log(spectr + eps)
return spectr
class Filterbank(torch.nn.Module):
"""computes filter bank (FBANK) features given spectral magnitudes.
Arguments
---------
n_mels : float
Number of Mel filters used to average the spectrogram.
log_mel : bool
If True, it computes the log of the FBANKs.
filter_shape : str
Shape of the filters ('triangular', 'rectangular', 'gaussian').
f_min : int
Lowest frequency for the Mel filters.
f_max : int
Highest frequency for the Mel filters.
n_fft : int
Number of fft points of the STFT. It defines the frequency resolution
(n_fft should be<= than win_len).
sample_rate : int
Sample rate of the input audio signal (e.g, 16000)
power_spectrogram : float
Exponent used for spectrogram computation.
amin : float
Minimum amplitude (used for numerical stability).
ref_value : float
Reference value used for the dB scale.
top_db : float
Minimum negative cut-off in decibels.
param_change_factor : bool
If freeze=False, this parameter affects the speed at which the filter
parameters (i.e., central_freqs and bands) can be changed. When high
(e.g., param_change_factor=1) the filters change a lot during training.
When low (e.g. param_change_factor=0.1) the filter parameters are more
stable during training
param_rand_factor : float
This parameter can be used to randomly change the filter parameters
(i.e, central frequencies and bands) during training. It is thus a
sort of regularization. param_rand_factor=0 does not affect, while
param_rand_factor=0.15 allows random variations within +-15% of the
standard values of the filter parameters (e.g., if the central freq
is 100 Hz, we can randomly change it from 85 Hz to 115 Hz).
freeze : bool
If False, it the central frequency and the band of each filter are
added into nn.parameters. If True, the standard frozen features
are computed.
Example
-------
>>> import torch
>>> compute_fbanks = Filterbank()
>>> inputs = torch.randn([10, 101, 201])
>>> features = compute_fbanks(inputs)
>>> features.shape
torch.Size([10, 101, 40])
"""
def __init__(
self,
n_mels=40,
log_mel=True,
filter_shape="triangular",
f_min=0,
f_max=8000,
n_fft=400,
sample_rate=16000,
power_spectrogram=2,
amin=1e-10,
ref_value=1.0,
top_db=80.0,
param_change_factor=1.0,
param_rand_factor=0.0,
freeze=True,
):
super().__init__()
self.n_mels = n_mels
self.log_mel = log_mel
self.filter_shape = filter_shape
self.f_min = f_min
self.f_max = f_max
self.n_fft = n_fft
self.sample_rate = sample_rate
self.power_spectrogram = power_spectrogram
self.amin = amin
self.ref_value = ref_value
self.top_db = top_db
self.freeze = freeze
self.n_stft = self.n_fft // 2 + 1
self.db_multiplier = math.log10(max(self.amin, self.ref_value))
self.device_inp = torch.device("cpu")
self.param_change_factor = param_change_factor
self.param_rand_factor = param_rand_factor
if self.power_spectrogram == 2:
self.multiplier = 10
else:
self.multiplier = 20
# Make sure f_min < f_max
if self.f_min >= self.f_max:
err_msg = "Require f_min: %f < f_max: %f" % (
self.f_min,
self.f_max,
)
logger.error(err_msg, exc_info=True)
# Filter definition
mel = torch.linspace(
self._to_mel(self.f_min), self._to_mel(self.f_max), self.n_mels + 2
)
hz = self._to_hz(mel)
# Computation of the filter bands
band = hz[1:] - hz[:-1]
self.band = band[:-1]
self.f_central = hz[1:-1]
# Adding the central frequency and the band to the list of nn param
if not self.freeze:
self.f_central = torch.nn.Parameter(
self.f_central / (self.sample_rate * self.param_change_factor)
)
self.band = torch.nn.Parameter(
self.band / (self.sample_rate * self.param_change_factor)
)
# Frequency axis
all_freqs = torch.linspace(0, self.sample_rate // 2, self.n_stft)
# Replicating for all the filters
self.all_freqs_mat = all_freqs.repeat(self.f_central.shape[0], 1)
def forward(self, spectrogram):
"""Returns the FBANks.
Arguments
---------
spectrogram : torch.Tensor
A batch of spectrogram tensors.
Returns
-------
fbanks : torch.Tensor
"""
# Computing central frequency and bandwidth of each filter
f_central_mat = self.f_central.repeat(
self.all_freqs_mat.shape[1], 1
).transpose(0, 1)
band_mat = self.band.repeat(self.all_freqs_mat.shape[1], 1).transpose(
0, 1
)
# Uncomment to print filter parameters
# print(self.f_central*self.sample_rate * self.param_change_factor)
# print(self.band*self.sample_rate* self.param_change_factor)
# Creation of the multiplication matrix. It is used to create
# the filters that average the computed spectrogram.
if not self.freeze:
f_central_mat = f_central_mat * (
self.sample_rate
* self.param_change_factor
* self.param_change_factor
)
band_mat = band_mat * (
self.sample_rate
* self.param_change_factor
* self.param_change_factor
)
# Regularization with random changes of filter central frequency and band
elif self.param_rand_factor != 0 and self.training:
rand_change = (
1.0
+ torch.rand(2) * 2 * self.param_rand_factor
- self.param_rand_factor
)
f_central_mat = f_central_mat * rand_change[0]
band_mat = band_mat * rand_change[1]
fbank_matrix = self._create_fbank_matrix(f_central_mat, band_mat).to(
spectrogram.device
)
sp_shape = spectrogram.shape
# Managing multi-channels case (batch, time, channels)
if len(sp_shape) == 4:
spectrogram = spectrogram.permute(0, 3, 1, 2)
spectrogram = spectrogram.reshape(
sp_shape[0] * sp_shape[3], sp_shape[1], sp_shape[2]
)
# FBANK computation
fbanks = torch.matmul(spectrogram, fbank_matrix)
if self.log_mel:
fbanks = self._amplitude_to_DB(fbanks)
# Reshaping in the case of multi-channel inputs
if len(sp_shape) == 4:
fb_shape = fbanks.shape
fbanks = fbanks.reshape(
sp_shape[0], sp_shape[3], fb_shape[1], fb_shape[2]
)
fbanks = fbanks.permute(0, 2, 3, 1)
return fbanks
@staticmethod
def _to_mel(hz):
"""Returns mel-frequency value corresponding to the input
frequency value in Hz.
Arguments
---------
hz : float
The frequency point in Hz.
Returns
-------
The mel-frequency value
"""
return 2595 * math.log10(1 + hz / 700)
@staticmethod
def _to_hz(mel):
"""Returns hz-frequency value corresponding to the input
mel-frequency value.
Arguments
---------
mel : float
The frequency point in the mel-scale.
Returns
-------
The hz-frequency value
"""
return 700 * (10 ** (mel / 2595) - 1)
def _triangular_filters(self, all_freqs, f_central, band):
"""Returns fbank matrix using triangular filters.
Arguments
---------
all_freqs : torch.Tensor
torch.Tensor gathering all the frequency points.
f_central : torch.Tensor
torch.Tensor gathering central frequencies of each filter.
band : torch.Tensor
torch.Tensor gathering the bands of each filter.
Returns
-------
fbank_matrix : torch.Tensor
"""
# Computing the slops of the filters
slope = (all_freqs - f_central) / band
left_side = slope + 1.0
right_side = -slope + 1.0
# Adding zeros for negative values
zero = torch.zeros(1, device=self.device_inp)
fbank_matrix = torch.max(
zero, torch.min(left_side, right_side)
).transpose(0, 1)
return fbank_matrix
def _rectangular_filters(self, all_freqs, f_central, band):
"""Returns fbank matrix using rectangular filters.
Arguments
---------
all_freqs : torch.Tensor
torch.Tensor gathering all the frequency points.
f_central : torch.Tensor
torch.Tensor gathering central frequencies of each filter.
band : torch.Tensor
torch.Tensor gathering the bands of each filter.
Returns
-------
fbank_matrix : torch.Tensor
"""
# cut-off frequencies of the filters
low_hz = f_central - band
high_hz = f_central + band
# Left/right parts of the filter
left_side = right_size = all_freqs.ge(low_hz)
right_size = all_freqs.le(high_hz)
fbank_matrix = (left_side * right_size).float().transpose(0, 1)
return fbank_matrix
def _gaussian_filters(
self, all_freqs, f_central, band, smooth_factor=torch.tensor(2)
):
"""Returns fbank matrix using gaussian filters.
Arguments
---------
all_freqs : torch.Tensor
torch.Tensor gathering all the frequency points.
f_central : torch.Tensor
torch.Tensor gathering central frequencies of each filter.
band : torch.Tensor
torch.Tensor gathering the bands of each filter.
smooth_factor: torch.Tensor
Smoothing factor of the gaussian filter. It can be used to employ
sharper or flatter filters.
Returns
-------
fbank_matrix : torch.Tensor
"""
fbank_matrix = torch.exp(
-0.5 * ((all_freqs - f_central) / (band / smooth_factor)) ** 2
).transpose(0, 1)
return fbank_matrix
def _create_fbank_matrix(self, f_central_mat, band_mat):
"""Returns fbank matrix to use for averaging the spectrum with
the set of filter-banks.
Arguments
---------
f_central_mat : torch.Tensor
torch.Tensor gathering central frequencies of each filter.
band_mat : torch.Tensor
torch.Tensor gathering the bands of each filter.
Returns
-------
fbank_matrix : torch.Tensor
"""
if self.filter_shape == "triangular":
fbank_matrix = self._triangular_filters(
self.all_freqs_mat, f_central_mat, band_mat
)
elif self.filter_shape == "rectangular":
fbank_matrix = self._rectangular_filters(
self.all_freqs_mat, f_central_mat, band_mat
)
else:
fbank_matrix = self._gaussian_filters(
self.all_freqs_mat, f_central_mat, band_mat
)
return fbank_matrix
def _amplitude_to_DB(self, x):
"""Converts linear-FBANKs to log-FBANKs.
Arguments
---------
x : torch.Tensor
A batch of linear FBANK tensors.
Returns
-------
x_db : torch.Tensor
"""
x_db = self.multiplier * torch.log10(torch.clamp(x, min=self.amin))
x_db -= self.multiplier * self.db_multiplier
# Setting up dB max. It is the max over time and frequency,
# Hence, of a whole sequence (sequence-dependent)
new_x_db_max = x_db.amax(dim=(-2, -1)) - self.top_db
# Clipping to dB max. The view is necessary as only a scalar is obtained
# per sequence.
x_db = torch.max(x_db, new_x_db_max.view(x_db.shape[0], 1, 1))
return x_db
class DCT(torch.nn.Module):
"""Computes the discrete cosine transform.
This class is primarily used to compute MFCC features of an audio signal
given a set of FBANK features as input.
Arguments
---------
input_size : int
Expected size of the last dimension in the input.
n_out : int
Number of output coefficients.
ortho_norm : bool
Whether to use orthogonal norm.
Example
-------
>>> import torch
>>> inputs = torch.randn([10, 101, 40])
>>> compute_mfccs = DCT(input_size=inputs.size(-1))
>>> features = compute_mfccs(inputs)
>>> features.shape
torch.Size([10, 101, 20])
"""
def __init__(self, input_size, n_out=20, ortho_norm=True):
super().__init__()
if n_out > input_size:
raise ValueError(
"Cannot select more DCT coefficients than inputs "
"(n_out=%i, n_in=%i)" % (n_out, input_size)
)
# Generate matrix for DCT transformation
n = torch.arange(float(input_size))
k = torch.arange(float(n_out)).unsqueeze(1)
dct = torch.cos(math.pi / float(input_size) * (n + 0.5) * k)
if ortho_norm:
dct[0] *= 1.0 / math.sqrt(2.0)
dct *= math.sqrt(2.0 / float(input_size))
else:
dct *= 2.0
self.dct_mat = dct.t()
def forward(self, x):
"""Returns the DCT of the input tensor.
Arguments
---------
x : torch.Tensor
A batch of tensors to transform, usually fbank features.
Returns
-------
dct : torch.Tensor
"""
# Managing multi-channels case
input_shape = x.shape
if len(input_shape) == 4:
x = x.reshape(x.shape[0] * x.shape[3], x.shape[1], x.shape[2])
# apply the DCT transform
dct = torch.matmul(x, self.dct_mat.to(x.device))
# Reshape in the case of multi-channels
if len(input_shape) == 4:
dct = dct.reshape(
input_shape[0], dct.shape[1], dct.shape[2], input_shape[3]
)
return dct
class Deltas(torch.nn.Module):
"""Computes delta coefficients (time derivatives).
Arguments
---------
input_size : int
The expected size of the inputs for parameter initialization.
window_length : int
Length of the window used to compute the time derivatives.
Example
-------
>>> inputs = torch.randn([10, 101, 20])
>>> compute_deltas = Deltas(input_size=inputs.size(-1))
>>> features = compute_deltas(inputs)
>>> features.shape
torch.Size([10, 101, 20])
"""
def __init__(self, input_size, window_length=5):
super().__init__()
self.n = (window_length - 1) // 2
self.denom = self.n * (self.n + 1) * (2 * self.n + 1) / 3
self.register_buffer(
"kernel",
torch.arange(
-self.n,
self.n + 1,
dtype=torch.float32,
).repeat(input_size, 1, 1),
)
def forward(self, x):
"""Returns the delta coefficients.
Arguments
---------
x : torch.Tensor
A batch of tensors.
Returns
-------
delta_coeff : torch.Tensor
"""
# Managing multi-channel deltas reshape tensor (batch*channel,time)
x = x.transpose(1, 2).transpose(2, -1)
or_shape = x.shape
if len(or_shape) == 4:
x = x.reshape(or_shape[0] * or_shape[2], or_shape[1], or_shape[3])
# Padding for time borders
x = torch.nn.functional.pad(x, (self.n, self.n), mode="replicate")
# Derivative estimation (with a fixed convolutional kernel)
delta_coeff = (
torch.nn.functional.conv1d(
x, self.kernel.to(x.device), groups=x.shape[1]
)
/ self.denom
)
# Retrieving the original dimensionality (for multi-channel case)
if len(or_shape) == 4:
delta_coeff = delta_coeff.reshape(
or_shape[0], or_shape[1], or_shape[2], or_shape[3]
)
delta_coeff = delta_coeff.transpose(1, -1).transpose(2, -1)
return delta_coeff
class ContextWindow(torch.nn.Module):
"""Computes the context window.
This class applies a context window by gathering multiple time steps
in a single feature vector. The operation is performed with a
convolutional layer based on a fixed kernel designed for that.
Arguments
---------
left_frames : int
Number of left frames (i.e, past frames) to collect.
right_frames : int
Number of right frames (i.e, future frames) to collect.
Example
-------
>>> import torch
>>> compute_cw = ContextWindow(left_frames=5, right_frames=5)
>>> inputs = torch.randn([10, 101, 20])
>>> features = compute_cw(inputs)
>>> features.shape
torch.Size([10, 101, 220])
"""
def __init__(self, left_frames=0, right_frames=0):
super().__init__()
self.left_frames = left_frames
self.right_frames = right_frames
self.context_len = self.left_frames + self.right_frames + 1
self.kernel_len = 2 * max(self.left_frames, self.right_frames) + 1
# Kernel definition
self.kernel = torch.eye(self.context_len, self.kernel_len)
if self.right_frames > self.left_frames:
lag = self.right_frames - self.left_frames
self.kernel = torch.roll(self.kernel, lag, 1)
self.first_call = True
def forward(self, x):
"""Returns the tensor with the surrounding context.
Arguments
---------
x : torch.Tensor
A batch of tensors.
Returns
-------
cw_x : torch.Tensor
The context-enriched tensor
"""
x = x.transpose(1, 2)
if self.first_call is True:
self.first_call = False
self.kernel = (
self.kernel.repeat(x.shape[1], 1, 1)
.view(x.shape[1] * self.context_len, self.kernel_len)
.unsqueeze(1)
)
# Managing multi-channel case
or_shape = x.shape
if len(or_shape) == 4:
x = x.reshape(or_shape[0] * or_shape[2], or_shape[1], or_shape[3])
# Compute context (using the estimated convolutional kernel)
cw_x = torch.nn.functional.conv1d(
x,
self.kernel.to(x.device),
groups=x.shape[1],
padding=max(self.left_frames, self.right_frames),
)
# Retrieving the original dimensionality (for multi-channel case)
if len(or_shape) == 4:
cw_x = cw_x.reshape(
or_shape[0], cw_x.shape[1], or_shape[2], cw_x.shape[-1]
)
cw_x = cw_x.transpose(1, 2)
return cw_x
def gaussian_statistics(
x: torch.Tensor,