-
Notifications
You must be signed in to change notification settings - Fork 9
/
train.py
250 lines (211 loc) · 10.7 KB
/
train.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
# -*- coding: utf-8 -*-
"""
Created on Wed Jul 24 19:29:31 2019
@author: a-kojima
Neural mask estimation for MVDR
this script suppots on-the-fly training for data-augmentation efficiently
"""
import numpy as np
import glob
from scipy import stats
import random
import soundfile as sf
import matplotlib.pyplot as pl
import sys
from maskestimator import model, shaper, feature, augment, util
from tensorflow.keras.callbacks import TensorBoard
#==========================================
# ANALYSIS PARAMETERS
#==========================================
SAMPLING_FREQUENCY = 16000
FFTL = 1024
SHIFT = 256
#==========================================
# ESURAL MASL ESTIMATOR PARAMETERS
#==========================================
LEFT_CONTEXT = 0
RIGHT_CONTEXT = 0
NUMBER_OF_SKIP_FRAME = 0
#==========================================
# NEURAL MASL ESTIMATOR TRAINNING PARAMERTERS
#==========================================
EPOCHS = 1
NUMBER_OF_UTTERANCE = 1 #15
TRUNCATE_GRAD = 7
LR = 0.001
SPEECH_DIRECTORY = r'./dataset/train/speech/*'
NOISE_DIRECTORY = r'./dataset/train/noise/*'
IS_DEBUG_SHOW_MASK_AND_SYNTHESIS = False
MODEL_NAME_PREFIX = r'./model/neaural_mask_estimator{}.hdf5'
LOSS_NAME_PREFIX = r'./model/neaural_mask_estimator.npy'
LOG_PATH = r'./tflog/'
RECURRENT_INIT = 0.04
VALIDATION_SPEC = r'./validation_features/val_spec.npy'
VALIDATION_SPEECH_MASK = r'./validation_features/noise_mask.npy'
VALIDATION_NOISE_MASK = r'./validation_features/speech_mask.npy'
NUMBER_OF_STACK = LEFT_CONTEXT + RIGHT_CONTEXT + 1
#==========================================
# augmentation parameters
#==========================================
'''
snr: [SNR20, SNR15, SNR10, SNR5, SNR0]
prob.: [0.2,... 0.2]
'''
SNR_generator = stats.rv_discrete(values=(np.array([0, 1, 2, 3, 4]),
(0.2, 0.2, 0.2, 0.2, 0.2)))
noise_list = glob.glob(NOISE_DIRECTORY)
RIR_CONVOLVE_CHANCE_RATE = 0.25 # ex., 0.5 means convolution speech with RIR with 50% chanve rate
#==========================================
# prepare speech and noise file list
#==========================================
speech_list = glob.glob(SPEECH_DIRECTORY)
noise_list = glob.glob(NOISE_DIRECTORY)
#==========================================
# get model
#==========================================
mask_estimator_generator = model.NeuralMaskEstimation(TRUNCATE_GRAD,
NUMBER_OF_STACK,
LR,
FFTL // 2 + 1,
recurrent_init=RECURRENT_INIT)
mask_estimator = mask_estimator_generator.get_model(is_stateful=False,
is_show_detail=True,
is_adapt=False)
#==========================================
# training data shaper
#==========================================
data_shaper = shaper.Shape_data(LEFT_CONTEXT,
RIGHT_CONTEXT,
TRUNCATE_GRAD,
NUMBER_OF_SKIP_FRAME )
#==========================================
# set tensorboard
#==========================================
callback = TensorBoard(LOG_PATH)
callback.set_model(mask_estimator)
#==========================================
# get features
#==========================================
feature_extractor = feature.Feature(SAMPLING_FREQUENCY, FFTL, SHIFT)
noise_generator = augment.Generate_random_noise(noise_list, SAMPLING_FREQUENCY)
reverbarent_generator = augment.RIR_convolve(SAMPLING_FREQUENCY)
#==========================================
# go training
#==========================================
utterance_count = 0
TRIM = np.int(0.05 * SAMPLING_FREQUENCY) # beginning and ending of uttearnce is not used for training
val_loss = np.array([])
test_loss = np.array([])
validate_features = np.load(VALIDATION_SPEC)
validate_label_sp = np.load(VALIDATION_SPEECH_MASK)
validate_label_n = np.load(VALIDATION_NOISE_MASK)
freq_grid = np.linspace(0, SAMPLING_FREQUENCY, FFTL)[0:FFTL // 2 + 1]
bin_index = np.argmin(np.abs(freq_grid - 2000))
for i in range(0, EPOCHS):
speech_list_shuffle = random.sample(speech_list, len(speech_list))
# go NN parameters optimizer
while True:
feature_stack = []
label_stack_sp = []
label_stack_n = []
if len(speech_list_shuffle) < NUMBER_OF_UTTERANCE:
break
# dumping frame until searching # of utterances
while True:
# all utterance is used for training
if len(speech_list_shuffle) <= NUMBER_OF_UTTERANCE:
break
index = np.random.randint(0, len(speech_list_shuffle) - 1, 1)[0]
audio_path = speech_list_shuffle[index]
speech_list_shuffle.pop(index) # remove uterance chosen yet
speech = sf.read(audio_path, dtype='float32')[0]
speech = speech[TRIM:-TRIM]
if len(speech) != 0:
speech = feature_extractor.add_white_noise(speech)
if IS_DEBUG_SHOW_MASK_AND_SYNTHESIS == True:
sf.write('./result/speech_clean.wav', speech , 16000)
SNR_index = SNR_generator.rvs(size=1)[0]
noise = noise_generator.get_noise(len(speech))
noise = feature_extractor.add_white_noise(noise)
if RIR_CONVOLVE_CHANCE_RATE != 0:
# convolve RIR
if np.random.randint(0, 1 // RIR_CONVOLVE_CHANCE_RATE, 1)[0] == 1:
speech, noise = reverbarent_generator.get_reverbant_speech(speech, noise)
snr_adjuster = augment.SNR_adjusting(speech, noise)
if SNR_index == 0:
SNR = 20
elif SNR_index == 1:
SNR = 15
elif SNR_index == 2:
SNR = 10
elif SNR_index == 3:
SNR = 5
elif SNR_index == 4:
SNR = 0
speech, noise = snr_adjuster.add_speech_to_noise(SNR)
speech, noise = snr_adjuster.avoid_clipping(speech, noise)
# if get mask after SNR adjusting
speech_spectrogram = feature_extractor.get_feature(speech)
noise_spectrogram = feature_extractor.get_feature(noise)
freq_grid = np.linspace(0, SAMPLING_FREQUENCY, FFTL)[0:FFTL // 2 + 1]
bin_index = np.argmin(np.abs(freq_grid - 2000))
speech_mask, noise_mask = feature_extractor.get_ideal_binary_mask_herman(speech_spectrogram,
noise_spectrogram,
threshold_bin=bin_index,
theta_sp_low=10**(-4),
theta_sp_high=10**(-5),
theta_n_low=10**(-5),#-0.01
theta_n_high=10**(-5)) #-0.02
noisy_spectrogram = (speech_spectrogram + noise_spectrogram)
noisy_spectrogram = (np.flipud(noisy_spectrogram))
speech_mask = np.flipud(speech_mask)
noise_mask = np.flipud(noise_mask)
noisy_spectrogram = feature_extractor.apply_cmvn(noisy_spectrogram)
noisy_spectrogram = noisy_spectrogram + np.random.normal(loc=0, scale=0.0001, size=np.shape(noisy_spectrogram))
features, label_sp, label_n = data_shaper.convert_for_train(noisy_spectrogram, speech_mask, noise_mask)
if len(features) != 0:
features = np.array(features)
label_sp = np.array(label_sp)
label_n = np.array(label_n)
if IS_DEBUG_SHOW_MASK_AND_SYNTHESIS == True:
sf.write('./result/speech_noisy.wav', speech + noise, 16000)
pl.figure(),
pl.imshow(noise_mask, aspect='auto', extent=[0, np.shape(noise_mask)[1], 0, 8000,])
pl.title('noise mask')
pl.figure(),
pl.imshow(speech_mask, aspect='auto',extent=[0, np.shape(noise_mask)[1], 0, 8000])
pl.title('sp mask')
pl.figure()
pl.imshow(noisy_spectrogram, aspect='auto')
pl.show()
sys.exit()
feature_stack.extend(features)
label_stack_sp.extend(label_sp)
label_stack_n.extend(label_n)
utterance_count = utterance_count + 1
if utterance_count == NUMBER_OF_UTTERANCE:
break
train_features = np.array(feature_stack)
train_label_sp = np.array(label_stack_sp)
train_label_n = np.array(label_stack_n)
if np.shape(train_features)[0] != 0:
shuffle_index = random.sample(range(0, np.shape(train_features)[0]), np.shape(train_features)[0])
history = mask_estimator.train_on_batch(x=train_features[shuffle_index, :, :],
y=[train_label_sp[shuffle_index, :], train_label_n[shuffle_index, :]])
print('train_loss:', history)
print('epoch:', i)
else:
break
# reset
feature_stack = []
label_stack = []
utterance_count = 0
# evaluate validation data
val_loss_r = mask_estimator.evaluate(x=validate_features,
y=[validate_label_sp, validate_label_n],
verbose=0)
print('val_loss:', val_loss_r)
util.write_log(callback, ['val_loss', 'speech_loss', 'noise_loss', 'accu_speech', 'accu_noise'], val_loss_r, i)
val_loss = np.append(val_loss, val_loss_r)
np.save(LOSS_NAME_PREFIX, val_loss)
mask_estimator.save_weights(MODEL_NAME_PREFIX.replace('{}', str(i)))