yangwang825
commited on
Commit
•
ca3e491
1
Parent(s):
8f84b12
Create helpers_svector.py
Browse files- helpers_svector.py +744 -0
helpers_svector.py
ADDED
@@ -0,0 +1,744 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
|
5 |
+
|
6 |
+
class Deltas(torch.nn.Module):
|
7 |
+
"""Computes delta coefficients (time derivatives).
|
8 |
+
|
9 |
+
Arguments
|
10 |
+
---------
|
11 |
+
win_length : int
|
12 |
+
Length of the window used to compute the time derivatives.
|
13 |
+
|
14 |
+
Example
|
15 |
+
-------
|
16 |
+
>>> inputs = torch.randn([10, 101, 20])
|
17 |
+
>>> compute_deltas = Deltas(input_size=inputs.size(-1))
|
18 |
+
>>> features = compute_deltas(inputs)
|
19 |
+
>>> features.shape
|
20 |
+
torch.Size([10, 101, 20])
|
21 |
+
"""
|
22 |
+
|
23 |
+
def __init__(
|
24 |
+
self, input_size, window_length=5,
|
25 |
+
):
|
26 |
+
super().__init__()
|
27 |
+
self.n = (window_length - 1) // 2
|
28 |
+
self.denom = self.n * (self.n + 1) * (2 * self.n + 1) / 3
|
29 |
+
|
30 |
+
self.register_buffer(
|
31 |
+
"kernel",
|
32 |
+
torch.arange(-self.n, self.n + 1, dtype=torch.float32,).repeat(
|
33 |
+
input_size, 1, 1
|
34 |
+
),
|
35 |
+
)
|
36 |
+
|
37 |
+
def forward(self, x):
|
38 |
+
"""Returns the delta coefficients.
|
39 |
+
|
40 |
+
Arguments
|
41 |
+
---------
|
42 |
+
x : tensor
|
43 |
+
A batch of tensors.
|
44 |
+
"""
|
45 |
+
# Managing multi-channel deltas reshape tensor (batch*channel,time)
|
46 |
+
x = x.transpose(1, 2).transpose(2, -1)
|
47 |
+
or_shape = x.shape
|
48 |
+
if len(or_shape) == 4:
|
49 |
+
x = x.reshape(or_shape[0] * or_shape[2], or_shape[1], or_shape[3])
|
50 |
+
|
51 |
+
# Padding for time borders
|
52 |
+
x = torch.nn.functional.pad(x, (self.n, self.n), mode="replicate")
|
53 |
+
|
54 |
+
# Derivative estimation (with a fixed convolutional kernel)
|
55 |
+
delta_coeff = (
|
56 |
+
torch.nn.functional.conv1d(
|
57 |
+
x, self.kernel.to(x.device), groups=x.shape[1]
|
58 |
+
)
|
59 |
+
/ self.denom
|
60 |
+
)
|
61 |
+
|
62 |
+
# Retrieving the original dimensionality (for multi-channel case)
|
63 |
+
if len(or_shape) == 4:
|
64 |
+
delta_coeff = delta_coeff.reshape(
|
65 |
+
or_shape[0], or_shape[1], or_shape[2], or_shape[3],
|
66 |
+
)
|
67 |
+
delta_coeff = delta_coeff.transpose(1, -1).transpose(2, -1)
|
68 |
+
|
69 |
+
return delta_coeff
|
70 |
+
|
71 |
+
|
72 |
+
class Filterbank(torch.nn.Module):
|
73 |
+
"""computes filter bank (FBANK) features given spectral magnitudes.
|
74 |
+
|
75 |
+
Arguments
|
76 |
+
---------
|
77 |
+
n_mels : float
|
78 |
+
Number of Mel filters used to average the spectrogram.
|
79 |
+
log_mel : bool
|
80 |
+
If True, it computes the log of the FBANKs.
|
81 |
+
filter_shape : str
|
82 |
+
Shape of the filters ('triangular', 'rectangular', 'gaussian').
|
83 |
+
f_min : int
|
84 |
+
Lowest frequency for the Mel filters.
|
85 |
+
f_max : int
|
86 |
+
Highest frequency for the Mel filters.
|
87 |
+
n_fft : int
|
88 |
+
Number of fft points of the STFT. It defines the frequency resolution
|
89 |
+
(n_fft should be<= than win_len).
|
90 |
+
sample_rate : int
|
91 |
+
Sample rate of the input audio signal (e.g, 16000)
|
92 |
+
power_spectrogram : float
|
93 |
+
Exponent used for spectrogram computation.
|
94 |
+
amin : float
|
95 |
+
Minimum amplitude (used for numerical stability).
|
96 |
+
ref_value : float
|
97 |
+
Reference value used for the dB scale.
|
98 |
+
top_db : float
|
99 |
+
Minimum negative cut-off in decibels.
|
100 |
+
freeze : bool
|
101 |
+
If False, it the central frequency and the band of each filter are
|
102 |
+
added into nn.parameters. If True, the standard frozen features
|
103 |
+
are computed.
|
104 |
+
param_change_factor: bool
|
105 |
+
If freeze=False, this parameter affects the speed at which the filter
|
106 |
+
parameters (i.e., central_freqs and bands) can be changed. When high
|
107 |
+
(e.g., param_change_factor=1) the filters change a lot during training.
|
108 |
+
When low (e.g. param_change_factor=0.1) the filter parameters are more
|
109 |
+
stable during training
|
110 |
+
param_rand_factor: float
|
111 |
+
This parameter can be used to randomly change the filter parameters
|
112 |
+
(i.e, central frequencies and bands) during training. It is thus a
|
113 |
+
sort of regularization. param_rand_factor=0 does not affect, while
|
114 |
+
param_rand_factor=0.15 allows random variations within +-15% of the
|
115 |
+
standard values of the filter parameters (e.g., if the central freq
|
116 |
+
is 100 Hz, we can randomly change it from 85 Hz to 115 Hz).
|
117 |
+
|
118 |
+
Example
|
119 |
+
-------
|
120 |
+
>>> import torch
|
121 |
+
>>> compute_fbanks = Filterbank()
|
122 |
+
>>> inputs = torch.randn([10, 101, 201])
|
123 |
+
>>> features = compute_fbanks(inputs)
|
124 |
+
>>> features.shape
|
125 |
+
torch.Size([10, 101, 40])
|
126 |
+
"""
|
127 |
+
|
128 |
+
def __init__(
|
129 |
+
self,
|
130 |
+
n_mels=40,
|
131 |
+
log_mel=True,
|
132 |
+
filter_shape="triangular",
|
133 |
+
f_min=0,
|
134 |
+
f_max=8000,
|
135 |
+
n_fft=400,
|
136 |
+
sample_rate=16000,
|
137 |
+
power_spectrogram=2,
|
138 |
+
amin=1e-10,
|
139 |
+
ref_value=1.0,
|
140 |
+
top_db=80.0,
|
141 |
+
param_change_factor=1.0,
|
142 |
+
param_rand_factor=0.0,
|
143 |
+
freeze=True,
|
144 |
+
):
|
145 |
+
super().__init__()
|
146 |
+
self.n_mels = n_mels
|
147 |
+
self.log_mel = log_mel
|
148 |
+
self.filter_shape = filter_shape
|
149 |
+
self.f_min = f_min
|
150 |
+
self.f_max = f_max
|
151 |
+
self.n_fft = n_fft
|
152 |
+
self.sample_rate = sample_rate
|
153 |
+
self.power_spectrogram = power_spectrogram
|
154 |
+
self.amin = amin
|
155 |
+
self.ref_value = ref_value
|
156 |
+
self.top_db = top_db
|
157 |
+
self.freeze = freeze
|
158 |
+
self.n_stft = self.n_fft // 2 + 1
|
159 |
+
self.db_multiplier = math.log10(max(self.amin, self.ref_value))
|
160 |
+
self.device_inp = torch.device("cpu")
|
161 |
+
self.param_change_factor = param_change_factor
|
162 |
+
self.param_rand_factor = param_rand_factor
|
163 |
+
|
164 |
+
if self.power_spectrogram == 2:
|
165 |
+
self.multiplier = 10
|
166 |
+
else:
|
167 |
+
self.multiplier = 20
|
168 |
+
|
169 |
+
# Make sure f_min < f_max
|
170 |
+
if self.f_min >= self.f_max:
|
171 |
+
err_msg = "Require f_min: %f < f_max: %f" % (
|
172 |
+
self.f_min,
|
173 |
+
self.f_max,
|
174 |
+
)
|
175 |
+
print(err_msg)
|
176 |
+
|
177 |
+
# Filter definition
|
178 |
+
mel = torch.linspace(
|
179 |
+
self._to_mel(self.f_min), self._to_mel(self.f_max), self.n_mels + 2
|
180 |
+
)
|
181 |
+
hz = self._to_hz(mel)
|
182 |
+
|
183 |
+
# Computation of the filter bands
|
184 |
+
band = hz[1:] - hz[:-1]
|
185 |
+
self.band = band[:-1]
|
186 |
+
self.f_central = hz[1:-1]
|
187 |
+
|
188 |
+
# Adding the central frequency and the band to the list of nn param
|
189 |
+
if not self.freeze:
|
190 |
+
self.f_central = torch.nn.Parameter(
|
191 |
+
self.f_central / (self.sample_rate * self.param_change_factor)
|
192 |
+
)
|
193 |
+
self.band = torch.nn.Parameter(
|
194 |
+
self.band / (self.sample_rate * self.param_change_factor)
|
195 |
+
)
|
196 |
+
|
197 |
+
# Frequency axis
|
198 |
+
all_freqs = torch.linspace(0, self.sample_rate // 2, self.n_stft)
|
199 |
+
|
200 |
+
# Replicating for all the filters
|
201 |
+
self.all_freqs_mat = all_freqs.repeat(self.f_central.shape[0], 1)
|
202 |
+
|
203 |
+
def forward(self, spectrogram):
|
204 |
+
"""Returns the FBANks.
|
205 |
+
|
206 |
+
Arguments
|
207 |
+
---------
|
208 |
+
x : tensor
|
209 |
+
A batch of spectrogram tensors.
|
210 |
+
"""
|
211 |
+
# Computing central frequency and bandwidth of each filter
|
212 |
+
f_central_mat = self.f_central.repeat(
|
213 |
+
self.all_freqs_mat.shape[1], 1
|
214 |
+
).transpose(0, 1)
|
215 |
+
band_mat = self.band.repeat(self.all_freqs_mat.shape[1], 1).transpose(
|
216 |
+
0, 1
|
217 |
+
)
|
218 |
+
|
219 |
+
# Uncomment to print filter parameters
|
220 |
+
# print(self.f_central*self.sample_rate * self.param_change_factor)
|
221 |
+
# print(self.band*self.sample_rate* self.param_change_factor)
|
222 |
+
|
223 |
+
# Creation of the multiplication matrix. It is used to create
|
224 |
+
# the filters that average the computed spectrogram.
|
225 |
+
if not self.freeze:
|
226 |
+
f_central_mat = f_central_mat * (
|
227 |
+
self.sample_rate
|
228 |
+
* self.param_change_factor
|
229 |
+
* self.param_change_factor
|
230 |
+
)
|
231 |
+
band_mat = band_mat * (
|
232 |
+
self.sample_rate
|
233 |
+
* self.param_change_factor
|
234 |
+
* self.param_change_factor
|
235 |
+
)
|
236 |
+
|
237 |
+
# Regularization with random changes of filter central frequency and band
|
238 |
+
elif self.param_rand_factor != 0 and self.training:
|
239 |
+
rand_change = (
|
240 |
+
1.0
|
241 |
+
+ torch.rand(2) * 2 * self.param_rand_factor
|
242 |
+
- self.param_rand_factor
|
243 |
+
)
|
244 |
+
f_central_mat = f_central_mat * rand_change[0]
|
245 |
+
band_mat = band_mat * rand_change[1]
|
246 |
+
|
247 |
+
fbank_matrix = self._create_fbank_matrix(f_central_mat, band_mat).to(
|
248 |
+
spectrogram.device
|
249 |
+
)
|
250 |
+
|
251 |
+
sp_shape = spectrogram.shape
|
252 |
+
|
253 |
+
# Managing multi-channels case (batch, time, channels)
|
254 |
+
if len(sp_shape) == 4:
|
255 |
+
spectrogram = spectrogram.permute(0, 3, 1, 2)
|
256 |
+
spectrogram = spectrogram.reshape(
|
257 |
+
sp_shape[0] * sp_shape[3], sp_shape[1], sp_shape[2]
|
258 |
+
)
|
259 |
+
|
260 |
+
# FBANK computation
|
261 |
+
fbanks = torch.matmul(spectrogram, fbank_matrix)
|
262 |
+
if self.log_mel:
|
263 |
+
fbanks = self._amplitude_to_DB(fbanks)
|
264 |
+
|
265 |
+
# Reshaping in the case of multi-channel inputs
|
266 |
+
if len(sp_shape) == 4:
|
267 |
+
fb_shape = fbanks.shape
|
268 |
+
fbanks = fbanks.reshape(
|
269 |
+
sp_shape[0], sp_shape[3], fb_shape[1], fb_shape[2]
|
270 |
+
)
|
271 |
+
fbanks = fbanks.permute(0, 2, 3, 1)
|
272 |
+
|
273 |
+
return fbanks
|
274 |
+
|
275 |
+
@staticmethod
|
276 |
+
def _to_mel(hz):
|
277 |
+
"""Returns mel-frequency value corresponding to the input
|
278 |
+
frequency value in Hz.
|
279 |
+
|
280 |
+
Arguments
|
281 |
+
---------
|
282 |
+
x : float
|
283 |
+
The frequency point in Hz.
|
284 |
+
"""
|
285 |
+
return 2595 * math.log10(1 + hz / 700)
|
286 |
+
|
287 |
+
@staticmethod
|
288 |
+
def _to_hz(mel):
|
289 |
+
"""Returns hz-frequency value corresponding to the input
|
290 |
+
mel-frequency value.
|
291 |
+
|
292 |
+
Arguments
|
293 |
+
---------
|
294 |
+
x : float
|
295 |
+
The frequency point in the mel-scale.
|
296 |
+
"""
|
297 |
+
return 700 * (10 ** (mel / 2595) - 1)
|
298 |
+
|
299 |
+
def _triangular_filters(self, all_freqs, f_central, band):
|
300 |
+
"""Returns fbank matrix using triangular filters.
|
301 |
+
|
302 |
+
Arguments
|
303 |
+
---------
|
304 |
+
all_freqs : Tensor
|
305 |
+
Tensor gathering all the frequency points.
|
306 |
+
f_central : Tensor
|
307 |
+
Tensor gathering central frequencies of each filter.
|
308 |
+
band : Tensor
|
309 |
+
Tensor gathering the bands of each filter.
|
310 |
+
"""
|
311 |
+
|
312 |
+
# Computing the slops of the filters
|
313 |
+
slope = (all_freqs - f_central) / band
|
314 |
+
left_side = slope + 1.0
|
315 |
+
right_side = -slope + 1.0
|
316 |
+
|
317 |
+
# Adding zeros for negative values
|
318 |
+
zero = torch.zeros(1, device=self.device_inp)
|
319 |
+
fbank_matrix = torch.max(
|
320 |
+
zero, torch.min(left_side, right_side)
|
321 |
+
).transpose(0, 1)
|
322 |
+
|
323 |
+
return fbank_matrix
|
324 |
+
|
325 |
+
def _rectangular_filters(self, all_freqs, f_central, band):
|
326 |
+
"""Returns fbank matrix using rectangular filters.
|
327 |
+
|
328 |
+
Arguments
|
329 |
+
---------
|
330 |
+
all_freqs : Tensor
|
331 |
+
Tensor gathering all the frequency points.
|
332 |
+
f_central : Tensor
|
333 |
+
Tensor gathering central frequencies of each filter.
|
334 |
+
band : Tensor
|
335 |
+
Tensor gathering the bands of each filter.
|
336 |
+
"""
|
337 |
+
|
338 |
+
# cut-off frequencies of the filters
|
339 |
+
low_hz = f_central - band
|
340 |
+
high_hz = f_central + band
|
341 |
+
|
342 |
+
# Left/right parts of the filter
|
343 |
+
left_side = right_size = all_freqs.ge(low_hz)
|
344 |
+
right_size = all_freqs.le(high_hz)
|
345 |
+
|
346 |
+
fbank_matrix = (left_side * right_size).float().transpose(0, 1)
|
347 |
+
|
348 |
+
return fbank_matrix
|
349 |
+
|
350 |
+
def _gaussian_filters(
|
351 |
+
self, all_freqs, f_central, band, smooth_factor=torch.tensor(2)
|
352 |
+
):
|
353 |
+
"""Returns fbank matrix using gaussian filters.
|
354 |
+
|
355 |
+
Arguments
|
356 |
+
---------
|
357 |
+
all_freqs : Tensor
|
358 |
+
Tensor gathering all the frequency points.
|
359 |
+
f_central : Tensor
|
360 |
+
Tensor gathering central frequencies of each filter.
|
361 |
+
band : Tensor
|
362 |
+
Tensor gathering the bands of each filter.
|
363 |
+
smooth_factor: Tensor
|
364 |
+
Smoothing factor of the gaussian filter. It can be used to employ
|
365 |
+
sharper or flatter filters.
|
366 |
+
"""
|
367 |
+
fbank_matrix = torch.exp(
|
368 |
+
-0.5 * ((all_freqs - f_central) / (band / smooth_factor)) ** 2
|
369 |
+
).transpose(0, 1)
|
370 |
+
|
371 |
+
return fbank_matrix
|
372 |
+
|
373 |
+
def _create_fbank_matrix(self, f_central_mat, band_mat):
|
374 |
+
"""Returns fbank matrix to use for averaging the spectrum with
|
375 |
+
the set of filter-banks.
|
376 |
+
|
377 |
+
Arguments
|
378 |
+
---------
|
379 |
+
f_central : Tensor
|
380 |
+
Tensor gathering central frequencies of each filter.
|
381 |
+
band : Tensor
|
382 |
+
Tensor gathering the bands of each filter.
|
383 |
+
smooth_factor: Tensor
|
384 |
+
Smoothing factor of the gaussian filter. It can be used to employ
|
385 |
+
sharper or flatter filters.
|
386 |
+
"""
|
387 |
+
if self.filter_shape == "triangular":
|
388 |
+
fbank_matrix = self._triangular_filters(
|
389 |
+
self.all_freqs_mat, f_central_mat, band_mat
|
390 |
+
)
|
391 |
+
|
392 |
+
elif self.filter_shape == "rectangular":
|
393 |
+
fbank_matrix = self._rectangular_filters(
|
394 |
+
self.all_freqs_mat, f_central_mat, band_mat
|
395 |
+
)
|
396 |
+
|
397 |
+
else:
|
398 |
+
fbank_matrix = self._gaussian_filters(
|
399 |
+
self.all_freqs_mat, f_central_mat, band_mat
|
400 |
+
)
|
401 |
+
|
402 |
+
return fbank_matrix
|
403 |
+
|
404 |
+
def _amplitude_to_DB(self, x):
|
405 |
+
"""Converts linear-FBANKs to log-FBANKs.
|
406 |
+
|
407 |
+
Arguments
|
408 |
+
---------
|
409 |
+
x : Tensor
|
410 |
+
A batch of linear FBANK tensors.
|
411 |
+
|
412 |
+
"""
|
413 |
+
|
414 |
+
x_db = self.multiplier * torch.log10(torch.clamp(x, min=self.amin))
|
415 |
+
x_db -= self.multiplier * self.db_multiplier
|
416 |
+
|
417 |
+
# Setting up dB max. It is the max over time and frequency,
|
418 |
+
# Hence, of a whole sequence (sequence-dependent)
|
419 |
+
new_x_db_max = x_db.amax(dim=(-2, -1)) - self.top_db
|
420 |
+
|
421 |
+
# Clipping to dB max. The view is necessary as only a scalar is obtained
|
422 |
+
# per sequence.
|
423 |
+
x_db = torch.max(x_db, new_x_db_max.view(x_db.shape[0], 1, 1))
|
424 |
+
|
425 |
+
return x_db
|
426 |
+
|
427 |
+
|
428 |
+
class STFT(torch.nn.Module):
|
429 |
+
"""computes the Short-Term Fourier Transform (STFT).
|
430 |
+
|
431 |
+
This class computes the Short-Term Fourier Transform of an audio signal.
|
432 |
+
It supports multi-channel audio inputs (batch, time, channels).
|
433 |
+
|
434 |
+
Arguments
|
435 |
+
---------
|
436 |
+
sample_rate : int
|
437 |
+
Sample rate of the input audio signal (e.g 16000).
|
438 |
+
win_length : float
|
439 |
+
Length (in ms) of the sliding window used to compute the STFT.
|
440 |
+
hop_length : float
|
441 |
+
Length (in ms) of the hope of the sliding window used to compute
|
442 |
+
the STFT.
|
443 |
+
n_fft : int
|
444 |
+
Number of fft point of the STFT. It defines the frequency resolution
|
445 |
+
(n_fft should be <= than win_len).
|
446 |
+
window_fn : function
|
447 |
+
A function that takes an integer (number of samples) and outputs a
|
448 |
+
tensor to be multiplied with each window before fft.
|
449 |
+
normalized_stft : bool
|
450 |
+
If True, the function returns the normalized STFT results,
|
451 |
+
i.e., multiplied by win_length^-0.5 (default is False).
|
452 |
+
center : bool
|
453 |
+
If True (default), the input will be padded on both sides so that the
|
454 |
+
t-th frame is centered at time t×hop_length. Otherwise, the t-th frame
|
455 |
+
begins at time t×hop_length.
|
456 |
+
pad_mode : str
|
457 |
+
It can be 'constant','reflect','replicate', 'circular', 'reflect'
|
458 |
+
(default). 'constant' pads the input tensor boundaries with a
|
459 |
+
constant value. 'reflect' pads the input tensor using the reflection
|
460 |
+
of the input boundary. 'replicate' pads the input tensor using
|
461 |
+
replication of the input boundary. 'circular' pads using circular
|
462 |
+
replication.
|
463 |
+
onesided : True
|
464 |
+
If True (default) only returns nfft/2 values. Note that the other
|
465 |
+
samples are redundant due to the Fourier transform conjugate symmetry.
|
466 |
+
|
467 |
+
Example
|
468 |
+
-------
|
469 |
+
>>> import torch
|
470 |
+
>>> compute_STFT = STFT(
|
471 |
+
... sample_rate=16000, win_length=25, hop_length=10, n_fft=400
|
472 |
+
... )
|
473 |
+
>>> inputs = torch.randn([10, 16000])
|
474 |
+
>>> features = compute_STFT(inputs)
|
475 |
+
>>> features.shape
|
476 |
+
torch.Size([10, 101, 201, 2])
|
477 |
+
"""
|
478 |
+
|
479 |
+
def __init__(
|
480 |
+
self,
|
481 |
+
sample_rate,
|
482 |
+
win_length=25,
|
483 |
+
hop_length=10,
|
484 |
+
n_fft=400,
|
485 |
+
window_fn=torch.hamming_window,
|
486 |
+
normalized_stft=False,
|
487 |
+
center=True,
|
488 |
+
pad_mode="constant",
|
489 |
+
onesided=True,
|
490 |
+
):
|
491 |
+
super().__init__()
|
492 |
+
self.sample_rate = sample_rate
|
493 |
+
self.win_length = win_length
|
494 |
+
self.hop_length = hop_length
|
495 |
+
self.n_fft = n_fft
|
496 |
+
self.normalized_stft = normalized_stft
|
497 |
+
self.center = center
|
498 |
+
self.pad_mode = pad_mode
|
499 |
+
self.onesided = onesided
|
500 |
+
|
501 |
+
# Convert win_length and hop_length from ms to samples
|
502 |
+
self.win_length = int(
|
503 |
+
round((self.sample_rate / 1000.0) * self.win_length)
|
504 |
+
)
|
505 |
+
self.hop_length = int(
|
506 |
+
round((self.sample_rate / 1000.0) * self.hop_length)
|
507 |
+
)
|
508 |
+
|
509 |
+
self.window = window_fn(self.win_length)
|
510 |
+
|
511 |
+
def forward(self, x):
|
512 |
+
"""Returns the STFT generated from the input waveforms.
|
513 |
+
|
514 |
+
Arguments
|
515 |
+
---------
|
516 |
+
x : tensor
|
517 |
+
A batch of audio signals to transform.
|
518 |
+
"""
|
519 |
+
|
520 |
+
# Managing multi-channel stft
|
521 |
+
or_shape = x.shape
|
522 |
+
if len(or_shape) == 3:
|
523 |
+
x = x.transpose(1, 2)
|
524 |
+
x = x.reshape(or_shape[0] * or_shape[2], or_shape[1])
|
525 |
+
|
526 |
+
stft = torch.stft(
|
527 |
+
x,
|
528 |
+
self.n_fft,
|
529 |
+
self.hop_length,
|
530 |
+
self.win_length,
|
531 |
+
self.window.to(x.device),
|
532 |
+
self.center,
|
533 |
+
self.pad_mode,
|
534 |
+
self.normalized_stft,
|
535 |
+
self.onesided,
|
536 |
+
return_complex=True,
|
537 |
+
)
|
538 |
+
|
539 |
+
stft = torch.view_as_real(stft)
|
540 |
+
|
541 |
+
# Retrieving the original dimensionality (batch,time, channels)
|
542 |
+
if len(or_shape) == 3:
|
543 |
+
stft = stft.reshape(
|
544 |
+
or_shape[0],
|
545 |
+
or_shape[2],
|
546 |
+
stft.shape[1],
|
547 |
+
stft.shape[2],
|
548 |
+
stft.shape[3],
|
549 |
+
)
|
550 |
+
stft = stft.permute(0, 3, 2, 4, 1)
|
551 |
+
else:
|
552 |
+
# (batch, time, channels)
|
553 |
+
stft = stft.transpose(2, 1)
|
554 |
+
|
555 |
+
return stft
|
556 |
+
|
557 |
+
|
558 |
+
def spectral_magnitude(
|
559 |
+
stft, power: int = 1, log: bool = False, eps: float = 1e-14
|
560 |
+
):
|
561 |
+
"""Returns the magnitude of a complex spectrogram.
|
562 |
+
|
563 |
+
Arguments
|
564 |
+
---------
|
565 |
+
stft : torch.Tensor
|
566 |
+
A tensor, output from the stft function.
|
567 |
+
power : int
|
568 |
+
What power to use in computing the magnitude.
|
569 |
+
Use power=1 for the power spectrogram.
|
570 |
+
Use power=0.5 for the magnitude spectrogram.
|
571 |
+
log : bool
|
572 |
+
Whether to apply log to the spectral features.
|
573 |
+
|
574 |
+
Example
|
575 |
+
-------
|
576 |
+
>>> a = torch.Tensor([[3, 4]])
|
577 |
+
>>> spectral_magnitude(a, power=0.5)
|
578 |
+
tensor([5.])
|
579 |
+
"""
|
580 |
+
spectr = stft.pow(2).sum(-1)
|
581 |
+
|
582 |
+
# Add eps avoids NaN when spectr is zero
|
583 |
+
if power < 1:
|
584 |
+
spectr = spectr + eps
|
585 |
+
spectr = spectr.pow(power)
|
586 |
+
|
587 |
+
if log:
|
588 |
+
return torch.log(spectr + eps)
|
589 |
+
return spectr
|
590 |
+
|
591 |
+
|
592 |
+
class ContextWindow(torch.nn.Module):
|
593 |
+
"""Computes the context window.
|
594 |
+
|
595 |
+
This class applies a context window by gathering multiple time steps
|
596 |
+
in a single feature vector. The operation is performed with a
|
597 |
+
convolutional layer based on a fixed kernel designed for that.
|
598 |
+
|
599 |
+
Arguments
|
600 |
+
---------
|
601 |
+
left_frames : int
|
602 |
+
Number of left frames (i.e, past frames) to collect.
|
603 |
+
right_frames : int
|
604 |
+
Number of right frames (i.e, future frames) to collect.
|
605 |
+
|
606 |
+
Example
|
607 |
+
-------
|
608 |
+
>>> import torch
|
609 |
+
>>> compute_cw = ContextWindow(left_frames=5, right_frames=5)
|
610 |
+
>>> inputs = torch.randn([10, 101, 20])
|
611 |
+
>>> features = compute_cw(inputs)
|
612 |
+
>>> features.shape
|
613 |
+
torch.Size([10, 101, 220])
|
614 |
+
"""
|
615 |
+
|
616 |
+
def __init__(
|
617 |
+
self, left_frames=0, right_frames=0,
|
618 |
+
):
|
619 |
+
super().__init__()
|
620 |
+
self.left_frames = left_frames
|
621 |
+
self.right_frames = right_frames
|
622 |
+
self.context_len = self.left_frames + self.right_frames + 1
|
623 |
+
self.kernel_len = 2 * max(self.left_frames, self.right_frames) + 1
|
624 |
+
|
625 |
+
# Kernel definition
|
626 |
+
self.kernel = torch.eye(self.context_len, self.kernel_len)
|
627 |
+
|
628 |
+
if self.right_frames > self.left_frames:
|
629 |
+
lag = self.right_frames - self.left_frames
|
630 |
+
self.kernel = torch.roll(self.kernel, lag, 1)
|
631 |
+
|
632 |
+
self.first_call = True
|
633 |
+
|
634 |
+
def forward(self, x):
|
635 |
+
"""Returns the tensor with the surrounding context.
|
636 |
+
|
637 |
+
Arguments
|
638 |
+
---------
|
639 |
+
x : tensor
|
640 |
+
A batch of tensors.
|
641 |
+
"""
|
642 |
+
|
643 |
+
x = x.transpose(1, 2)
|
644 |
+
|
645 |
+
if self.first_call is True:
|
646 |
+
self.first_call = False
|
647 |
+
self.kernel = (
|
648 |
+
self.kernel.repeat(x.shape[1], 1, 1)
|
649 |
+
.view(x.shape[1] * self.context_len, self.kernel_len,)
|
650 |
+
.unsqueeze(1)
|
651 |
+
)
|
652 |
+
|
653 |
+
# Managing multi-channel case
|
654 |
+
or_shape = x.shape
|
655 |
+
if len(or_shape) == 4:
|
656 |
+
x = x.reshape(or_shape[0] * or_shape[2], or_shape[1], or_shape[3])
|
657 |
+
|
658 |
+
# Compute context (using the estimated convolutional kernel)
|
659 |
+
cw_x = torch.nn.functional.conv1d(
|
660 |
+
x,
|
661 |
+
self.kernel.to(x.device),
|
662 |
+
groups=x.shape[1],
|
663 |
+
padding=max(self.left_frames, self.right_frames),
|
664 |
+
)
|
665 |
+
|
666 |
+
# Retrieving the original dimensionality (for multi-channel case)
|
667 |
+
if len(or_shape) == 4:
|
668 |
+
cw_x = cw_x.reshape(
|
669 |
+
or_shape[0], cw_x.shape[1], or_shape[2], cw_x.shape[-1]
|
670 |
+
)
|
671 |
+
|
672 |
+
cw_x = cw_x.transpose(1, 2)
|
673 |
+
|
674 |
+
return cw_x
|
675 |
+
|
676 |
+
|
677 |
+
class Fbank(torch.nn.Module):
|
678 |
+
|
679 |
+
def __init__(
|
680 |
+
self,
|
681 |
+
deltas=False,
|
682 |
+
context=False,
|
683 |
+
requires_grad=False,
|
684 |
+
sample_rate=16000,
|
685 |
+
f_min=0,
|
686 |
+
f_max=None,
|
687 |
+
n_fft=400,
|
688 |
+
n_mels=40,
|
689 |
+
filter_shape="triangular",
|
690 |
+
param_change_factor=1.0,
|
691 |
+
param_rand_factor=0.0,
|
692 |
+
left_frames=5,
|
693 |
+
right_frames=5,
|
694 |
+
win_length=25,
|
695 |
+
hop_length=10,
|
696 |
+
):
|
697 |
+
super().__init__()
|
698 |
+
self.deltas = deltas
|
699 |
+
self.context = context
|
700 |
+
self.requires_grad = requires_grad
|
701 |
+
|
702 |
+
if f_max is None:
|
703 |
+
f_max = sample_rate / 2
|
704 |
+
|
705 |
+
self.compute_STFT = STFT(
|
706 |
+
sample_rate=sample_rate,
|
707 |
+
n_fft=n_fft,
|
708 |
+
win_length=win_length,
|
709 |
+
hop_length=hop_length,
|
710 |
+
)
|
711 |
+
self.compute_fbanks = Filterbank(
|
712 |
+
sample_rate=sample_rate,
|
713 |
+
n_fft=n_fft,
|
714 |
+
n_mels=n_mels,
|
715 |
+
f_min=f_min,
|
716 |
+
f_max=f_max,
|
717 |
+
freeze=not requires_grad,
|
718 |
+
filter_shape=filter_shape,
|
719 |
+
param_change_factor=param_change_factor,
|
720 |
+
param_rand_factor=param_rand_factor,
|
721 |
+
)
|
722 |
+
self.compute_deltas = Deltas(input_size=n_mels)
|
723 |
+
self.context_window = ContextWindow(
|
724 |
+
left_frames=left_frames, right_frames=right_frames,
|
725 |
+
)
|
726 |
+
|
727 |
+
def forward(self, wav):
|
728 |
+
"""Returns a set of features generated from the input waveforms.
|
729 |
+
|
730 |
+
Arguments
|
731 |
+
---------
|
732 |
+
wav : tensor
|
733 |
+
A batch of audio signals to transform to features.
|
734 |
+
"""
|
735 |
+
STFT = self.compute_STFT(wav)
|
736 |
+
mag = spectral_magnitude(STFT)
|
737 |
+
fbanks = self.compute_fbanks(mag)
|
738 |
+
if self.deltas:
|
739 |
+
delta1 = self.compute_deltas(fbanks)
|
740 |
+
delta2 = self.compute_deltas(delta1)
|
741 |
+
fbanks = torch.cat([fbanks, delta1, delta2], dim=2)
|
742 |
+
if self.context:
|
743 |
+
fbanks = self.context_window(fbanks)
|
744 |
+
return fbanks
|