-
Notifications
You must be signed in to change notification settings - Fork 21
/
utils.py
281 lines (224 loc) · 8.06 KB
/
utils.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
#!pip install soundfile
#!pip install librosa==0.8.0
#!pip install scipy==1.5.2
# use 3 lines above in console or Google Colaboratory (colab)
import numpy as np
import matplotlib.pyplot as plt
import librosa
import librosa.display as display
import librosa.feature
import soundfile as sf
from scipy.signal import resample
import math
import torch
##################################
# audio generation utils
##################################
def extract_audio(filename):
"""
Extract audio given the filename (.wav, .flac, etc format)
"""
audio, rate = sf.read(filename, always_2d=True)
audio = np.reshape(audio, (1, -1))
audio = audio[0]
time = np.linspace(0, len(audio)/rate, len(audio), endpoint=False)
return audio, time, rate
def generate_spec(audio_sequence, rate, n_fft=2048, hop_length=512):
"""
Generate spectrogram using librosa
audio_sequence: list representing waveform
rate: sampling rate (16000 for all LibriSpeech audios)
nfft and hop_length: stft parameters
"""
S = librosa.feature.melspectrogram(audio_sequence, sr=rate, n_fft=n_fft, hop_length=hop_length, n_mels=128, fmin=20,
fmax=8300)
log_spectra = librosa.power_to_db(S, ref=np.mean, top_db=80)
return log_spectra
def reconstruct_wave(spec, rate=16000, normalize_data=False):
"""
Reconstruct waveform
spec: spectrogram generated using Librosa
rate: sampling rate
"""
power = librosa.db_to_power(spec, ref=5.0)
audio = librosa.feature.inverse.mel_to_audio(power, sr=rate, n_fft=2048, hop_length=512)
out_audio = audio / np.max(audio) if normalize_data else audio
return out_audio
def normalize(spec, eps=1e-6):
"""
Normalize spectrogram with zero mean and unitary variance
spec: spectrogram generated using Librosa
"""
mean = spec.mean()
std = spec.std()
spec_norm = (spec - mean) / (std + eps)
return spec_norm, (mean, std)
def minmax_scaler(spec):
"""
min max scaler over spectrogram
"""
spec_max = np.max(spec)
spec_min = np.min(spec)
return (spec-spec_min)/(spec_max - spec_min), (spec_max, spec_min)
def linear_scaler(spec):
"""
linear scaler over spectrogram
min value -> -1 and max value -> 1
"""
spec_max = np.max(spec)
spec_min = np.min(spec)
m = 2/(spec_max-spec_min)
n = (spec_max + spec_min)/(spec_min-spec_max)
return m*spec + n, (m, n)
def split_specgram(example, clean_example, frames = 11):
"""
Split specgram in groups of frames, the purpose is prepare data for the LSTM model input
example: reverberant spectrogram
clean_example: clean or target spectrogram
return data input to the LSTM model and targets
"""
clean_spec = clean_example[0, :, :]
rev_spec = example[0, :, :]
n, m = clean_spec.shape
targets = torch.zeros((m-frames+1, n))
data = torch.zeros((m-frames+1, n*frames))
idx_target = frames//2
for i in range(m-frames+1):
try:
targets[i, :] = clean_spec[:, idx_target]
data[i, :] = torch.reshape(rev_spec[:, i:i+frames], (1, -1))[0, :]
idx_target += 1
except (IndexError):
pass
return data, targets
def split_realdata(example, frames = 11):
"""
Split 1 specgram in groups of frames, the purpose is prepare data for the LSTM and MLP model input
example: reverberant ''real'' (not simulated) spectrogram
return data input to the LSTM or MLP model
"""
rev_spec = example[0, :, :]
n, m = rev_spec.shape
data = torch.zeros((m-frames+1, n*frames))
for i in range(m-frames+1):
data[i, :] = torch.reshape(rev_spec[:, i:i+frames], (1, -1))[0, :]
return data
def prepare_data(X, y, display = False):
"""
Use split_specgram to split all specgrams
X: tensor containing reverberant spectrograms
y: tensor containing target spectrograms
"""
data0, target0 = split_specgram(X[0, :, :, :], y[0, :, :, :])
total_data = data0.cuda()
targets = target0.cuda()
for i in range(1, X.shape[0]):
if display:
print("Specgram n°" + str(i))
data_i, target_i = split_specgram(X[i, :, :, :], y[i, :, :, :])
total_data = torch.cat((total_data, data_i.cuda()), 0)
targets = torch.cat((targets, target_i.cuda()), 0)
return total_data, targets
def split_for_supression(rev_tensor, target_tensor):
"""
Given reverberant and target tensor with shape (#examples, 1, 128, 340)
return tensors with the same information, but with shape (#examples*340, 128)
"""
rev_transform = torch.tensor([])
target_transform = torch.tensor([])
for example in range(rev_tensor.shape[0]):
rev_transform = torch.cat((rev_transform, rev_tensor[example, 0, :, :].T))
if (target_tensor!=None):
for example in range(target_tensor.shape[0]):
target_transform = torch.cat((target_transform, target_tensor[example, 0, :, :].T))
return rev_transform, target_transform
def normalize_per_frame(spec_transpose):
"""
Normalize over spectrogram rows
"""
means = []
stds = []
norm_spec = torch.zeros(spec_transpose.shape)
for spec_row in range(norm_spec.shape[0]):
current_mean = spec_transpose[spec_row, :].mean()
current_std = spec_transpose[spec_row, :].std()
means.append(current_mean)
stds.append(current_std)
norm_spec[spec_row, :] = (spec_transpose[spec_row, :]- current_mean)/(current_std+1e-6)
return norm_spec, (means, stds)
def denormalize_per_frame(norm_spec_transpose, means, stds):
"""
denormalize row by row using means and stds given by normalize_per_frame
"""
denorm_spec = torch.zeros(norm_spec_transpose.shape)
for spec_row in range(norm_spec_transpose.shape[0]):
denorm_spec[spec_row, :] = (norm_spec_transpose[spec_row, :])*(stds[spec_row] + 1e-6) + means[spec_row]
return denorm_spec.T
#################################
# reverberation utils
#################################
def zero_pad(x, k):
"""
add k zeros to x signal
"""
return np.append(x, np.zeros(k))
def awgn(signal, regsnr):
"""
add random noise to signal
regsnr: signal to noise ratio
"""
sigpower = sum([math.pow(abs(signal[i]), 2) for i in range(len(signal))])
sigpower = sigpower / len(signal)
noisepower = sigpower / (math.pow(10, regsnr / 10))
sample = np.random.normal(0, 1, len(signal))
noise = math.sqrt(noisepower) * sample
return noise
def discrete_conv(x, h, x_fs, h_fs, snr=30, aug_factor=1):
"""
Convolution using fft
x: speech waveform
h: RIR waveform
x_fs: speech signal sampling rate (if is not 16000 the signal will be resampled)
h_fs: RIR signal sampling rate (if is not 16000 the signal will be resampled)
Based on https://github.com/vtolani95/convolution/blob/master/reverb.py
"""
numSamples_h = round(len(h) / h_fs * 16000)
numSamples_x = round(len(x) / x_fs * 16000)
if h_fs != 16000:
h = resample(h, numSamples_h) # resample RIR
if x_fs != 16000:
x = resample(x, numSamples_x) # resample speech signal
L, P = len(x), len(h)
h_zp = zero_pad(h, L - 1)
x_zp = zero_pad(x, P - 1)
X = np.fft.fft(x_zp)
output = np.fft.ifft(X * np.fft.fft(h_zp)).real
output = aug_factor * output + x_zp
output = output + awgn(output, snr)
return output
###################################
#plot utils
###################################
def graph_spec(spec, rate=16000, title=False):
"""
plot spectrogram
spec: spectrogram generated using Librosa
rate: sampling rate
"""
plt.figure()
display.specshow(spec, sr=rate, y_axis='mel', x_axis='time')
plt.colorbar(format='%+2.0f dB')
if (title):
plt.title('Log-Power spectrogram')
plt.tight_layout()
def plot_time_wave(audio, rate=16000):
"""
plot waveform given speech audio
audio: array containing waveform
rate: sampling rate
"""
time = np.linspace(0, len(audio)/rate, len(audio), endpoint=False)
plt.figure()
plt.plot(time, audio)
plt.xlabel("Time (secs)")
plt.ylabel("Power")