-
Notifications
You must be signed in to change notification settings - Fork 0
/
preprocessing.py
377 lines (249 loc) · 10.7 KB
/
preprocessing.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
import os
import shutil
from glob import glob
import random
import tensorflow as tf
import tensorflow_io as tfio
import matplotlib.pyplot as plt
@tf.function
def irm(background, speech, beta=0.5):
background = tf.cast(background, tf.float32)
speech = tf.cast(speech, tf.float32)
sq_s = tf.math.square(speech)
sq_b = tf.math.square(background)
s_b = tf.math.add(sq_s, sq_b)
mask = tf.divide(sq_s, s_b)
return tf.pow(mask, beta)
@tf.function
def load_wav(file_path):
file_contents = tf.io.read_file(file_path)
sig, sr = tf.audio.decode_wav(file_contents, desired_channels=1)
sig = tf.squeeze(sig, axis=-1)
return sig, sr
@tf.function
def load_stereo_wav(file_path):
file_contents = tf.io.read_file(file_path)
sig, sr = tf.audio.decode_wav(file_contents, desired_channels=2)
return sig, sr
@tf.function
def save_wav(file_path, signal, sr):
signal = tf.expand_dims(signal, -1)
encoded_signal = tf.audio.encode_wav(signal, sample_rate=sr)
tf.io.write_file(file_path, encoded_signal)
return
@tf.function
def save_stereo_wav(file_path, signal, sr):
encoded_signal = tf.audio.encode_wav(signal, sample_rate=sr)
tf.io.write_file(file_path, encoded_signal)
return
@tf.function
def istft(stft):
frame_length = 320
frame_step = 160
i_s = tf.signal.inverse_stft(
stft,
frame_length,
frame_step,
fft_length=frame_length,
window_fn=tf.signal.inverse_stft_window_fn(frame_step)
)
return i_s
@tf.function
def ibm(ref_z, mixed_z):
snr = tf.divide(tf.abs(ref_z), tf.abs(mixed_z))
mask2 = tf.math.round(snr)
mask2 = tf.where(tf.math.is_nan(mask2), 1.0, mask2)
mask2 = tf.where(mask2 > 1.0, 1.0, mask2)
mask1 = 1.0 - mask2
return mask1, mask2
def test_synthesis(lpb_path, speech_path, out_wav_path, seconds, target):
signal_len = seconds * 16000
mic_sig, mic_sr = load_wav(speech_path)[:signal_len]
ref_sig, ref_sr = load_wav(lpb_path)[:signal_len]
mixed_signal = tf.add(ref_sig, mic_sig)
f_len = 320
f_step = 160
z1 = tf.signal.stft(mic_sig, frame_length=f_len, frame_step=f_step, fft_length=f_len)
z2 = tf.signal.stft(ref_sig, frame_length=f_len, frame_step=f_step, fft_length=f_len)
z3 = tf.signal.stft(mixed_signal, frame_length=f_len, frame_step=f_step, fft_length=f_len)
if target == "ibm":
speech_mask, ref_mask = ibm(z2, z3)
z_masked_2 = tf.multiply(z3, tf.cast(ref_mask, tf.complex64))
rec_music = istft(z_masked_2)
elif target == "irm":
speech_mask = irm(z2, z1)
else:
print(f"Target mask {target} not supported.")
return
z_masked_1 = tf.multiply(z3, tf.cast(speech_mask, tf.complex64))
rec_speech = istft(z_masked_1)
f, ax = plt.subplots(2, 1)
ax[0].set_title("Mixed")
ax[0].plot(mixed_signal)
ax[1].set_title("Reconstructed Speech")
ax[1].plot(rec_speech)
f.tight_layout()
plt.show()
save_wav(out_wav_path, rec_speech, 16000)
return
def split_ref_and_speech_data(input_path, output_path, valid_split):
assert valid_split < 1.0, "Percentage of validation data must be less than 1.0"
ref_path = os.path.join(input_path, "REF")
speech_path = os.path.join(input_path, "SPEECH")
split_files(ref_path, output_path, valid_split)
split_files(speech_path, output_path, valid_split)
return
def split_files(input_path, output_path, valid_split):
data_type = os.path.basename(input_path)
all_files = glob(os.path.join(input_path, "*.wav"))
num_files = len(all_files)
random.shuffle(all_files)
valid_idx = int(valid_split * num_files)
valid_files = all_files[:valid_idx]
train_files = all_files[valid_idx:]
move_files(train_files, "TRAIN", data_type, output_path)
move_files(valid_files, "VALID", data_type, output_path)
return
def move_files(files, dataset, data_type, output_path):
dataset_output_path = os.path.join(output_path, dataset, data_type)
if not os.path.exists(dataset_output_path):
os.makedirs(dataset_output_path)
for file in files:
dst = os.path.join(dataset_output_path, os.path.basename(file))
shutil.copyfile(file, dst)
return
def split_wavs_into_1s_files(input_path, output_path):
datasets = os.listdir(input_path)
for dataset in datasets:
dataset_path = os.path.join(input_path, dataset)
signal_types = os.listdir(dataset_path)
for signal_type in signal_types:
wav_files = glob(os.path.join(dataset_path, signal_type, "*.wav"))
out_dir = os.path.join(output_path, dataset, signal_type)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
for wav in wav_files:
sig, sr = load_wav(tf.constant(wav))
len_sig = len(sig)
end = len_sig - len_sig % sr
cut_sig = sig[:end]
num_splits = int(len(cut_sig) / sr)
if num_splits == 0:
print(f"Not enough splits ({num_splits}) in {wav}. SKIPPING")
continue
one_second_sigs = tf.split(cut_sig, num_splits)
for i, one_second_sig in enumerate(one_second_sigs):
file_name = f"{os.path.splitext(os.path.basename(wav))[0]}_{i}.wav"
dst = os.path.join(out_dir, file_name)
save_wav(tf.constant(dst), one_second_sig, sr)
print(f"Done splitting {dataset} - > {signal_type} -> {wav}")
return
def combine_speech_and_ref(input_path, output_path):
datasets = os.listdir(input_path)
output_qty_dict = {"TRAIN": 1_000_000, "VALID": 100_000}
for dataset in datasets:
dataset_path = os.path.join(input_path, dataset)
speech_dir = os.path.join(dataset_path, "SPEECH")
ref_dir = os.path.join(dataset_path, "REF")
speech_files = glob(os.path.join(speech_dir, "*.wav"))
ref_files = glob(os.path.join(ref_dir, "*.wav"))
random.shuffle(speech_files)
random.shuffle(ref_files)
out_dir = os.path.join(output_path, dataset)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
num_output = output_qty_dict[dataset]
num_speech_files = len(speech_files)
num_to_mix = num_output // num_speech_files
for i, speech_file in enumerate(speech_files):
speech_sig, _ = load_wav(tf.constant(speech_file))
speech_name = os.path.splitext(os.path.basename(speech_file))[0]
for j, ref_file in enumerate(random.sample(ref_files, num_to_mix)):
ref_name = os.path.splitext(os.path.basename(ref_file))[0]
ref_sig, sr = load_wav(tf.constant(ref_file))
mixed_sig = speech_sig + ref_sig
stereo_sig = tf.stack((mixed_sig, ref_sig), axis=-1)
save_name = f"{speech_name}_SS_{ref_name}_RS_{i}_{j}.wav"
dst = os.path.join(out_dir, save_name)
save_stereo_wav(tf.constant(dst), stereo_sig, sr)
return
def _bytes_feature(value):
"""Returns a bytes_list from a string / byte."""
if isinstance(value, type(tf.constant(0))):
value = value.numpy() # BytesList won't unpack a string from an EagerTensor.
return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))
def get_dataset(filenames, batch_size=64):
dataset = tf.data.TFRecordDataset(filenames)
dataset = dataset.map(parse_spect_record, num_parallel_calls=tf.data.AUTOTUNE)
dataset = dataset.batch(batch_size, num_parallel_calls=tf.data.AUTOTUNE)
dataset = dataset.prefetch(tf.data.AUTOTUNE)
return dataset
def aec_example(spectrogram, target_mask):
"""Return TF example protobuf containing spectrogram and target mask."""
serialized_spect = tf.io.serialize_tensor(spectrogram)
serialized_mask = tf.io.serialize_tensor(target_mask)
spect_feat_bytes = _bytes_feature(serialized_spect)
mask_feat_bytes = _bytes_feature(serialized_mask)
feature = {
'spectrogram': spect_feat_bytes,
'target_mask': mask_feat_bytes
}
example_proto = tf.train.Example(
features=tf.train.Features(
feature=feature
)
)
return example_proto
def parse_spect_record(example):
"""Parses one example given feature description."""
feature_description = {
'spectrogram': tf.io.FixedLenFeature([], tf.string),
'target_mask': tf.io.FixedLenFeature([], tf.string)
}
content = tf.io.parse_single_example(example, feature_description)
spect = content['spectrogram']
target_mask = content['target_mask']
spect = tf.io.parse_tensor(spect, tf.complex64)
target_mask = tf.io.parse_tensor(target_mask, tf.float32)
return spect, target_mask
def write_examples_to_record(examples, output_path):
"""Write examples to tfrecords file until file reaches specified size."""
tfrecord_file_size = 10.2 ** 8
counter = 0
tfrecord_save_path = os.path.join(output_path, f"{counter}.tfrecords")
writer = tf.io.TFRecordWriter(tfrecord_save_path)
for example in examples:
if os.path.getsize(tfrecord_save_path) > tfrecord_file_size:
counter += 1
writer.close()
tfrecord_save_path = os.path.join(output_path, f"{counter}.tfrecords")
writer = tf.io.TFRecordWriter(tfrecord_save_path)
writer.write(example.SerializeToString())
writer.close()
return
def generate_example(stereo_file):
spectrogram, target_mask = extract_features(tf.constant(stereo_file))
example = aec_example(spectrogram, target_mask)
return example
def generate_spectrogram_and_mask_tfrecords(input_path, output_path):
datasets = os.listdir(input_path)
for dataset in datasets:
dataset_output_path = os.path.join(output_path, dataset)
if not os.path.exists(dataset_output_path):
os.mkdir(dataset_output_path)
dataset_input_path = os.path.join(input_path, dataset)
stereo_files = glob(os.path.join(dataset_input_path, "*.wav"))
examples = map(generate_example, stereo_files)
write_examples_to_record(examples, dataset_output_path)
return
@tf.function
def extract_features(stereo_wav_file_path):
sigs, sr = load_stereo_wav(stereo_wav_file_path)
duration = 4000 # 250ms
mixed_sig = sigs[:, 0][:duration]
ref_sig = sigs[:, 1][:duration]
z_mixed = tf.signal.stft(mixed_sig, frame_length=320, frame_step=160, fft_length=320)
z_ref = tf.signal.stft(ref_sig, frame_length=320, frame_step=160, fft_length=320)
target_mask, _ = ibm(z_ref, z_mixed) # (24, 161)
f_m = tf.concat([z_mixed, z_ref], axis=-1) # (24, 322)
return f_m, target_mask