-
Notifications
You must be signed in to change notification settings - Fork 59
/
main_DNN_TF_Masking.py
181 lines (153 loc) · 8.75 KB
/
main_DNN_TF_Masking.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
# -*- coding: utf-8 -*-
"""
Created on Wed Mar 28 11:12:48 2018
@author: eesungkim
"""
import os
import librosa
import argparse
import numpy as np
import scipy.io.wavfile as wav
import librosa.display
from utils.utils import *
from utils.estnoise_ms import *
import tensorflow as tf
from keras.layers import Input
from keras.models import Model
from keras import backend as k
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten, Conv2D, MaxPooling2D, BatchNormalization, Activation, AveragePooling2D
import keras
from keras import metrics
from keras.layers.normalization import BatchNormalization
from keras import initializers
from keras.layers import LeakyReLU, PReLU, ELU
def estimate_IRM(magnitude_clean,magnitude_noise):
denominator = np.abs(magnitude_clean) ** 2 + np.abs(magnitude_noise) ** 2
return (np.abs(magnitude_clean) ** 2 / denominator) ** (0.5)
def DNN_TF_Masking(args):
"""Speech Enhancement using DNN TF Masking"""
PATH_ROOT = os.getcwd()
os.chdir(PATH_ROOT)
# noisy_train ; input of DNN
path_dnn_noisy_train = os.path.join(PATH_ROOT, args.input_noisy_train)
dnn_magnitude_noisy_train,_,sr = perform_stft(path_dnn_noisy_train, args)
dnn_magnitude_noisy_train = apply_context_window(dnn_magnitude_noisy_train.T, args.left_context, args.right_context).T
# clean_train ; output of DNN
path_dnn_clean_train = os.path.join(PATH_ROOT, args.input_clean_train)
dnn_magnitude_clean_train,_,_ = perform_stft(path_dnn_clean_train, args)
# noise_train
path_noise = os.path.join(PATH_ROOT, args.input_noise)
dnn_magnitude_noise_train,_,_ = perform_stft(path_noise, args)
path_clean_test = os.path.join(PATH_ROOT , args.input_clean_test)
(sr, clean_test) = wav.read(path_clean_test)
# noisy_test
path_noisy_test = os.path.join(PATH_ROOT, args.input_noisy_test)
(sr, noisy_test) = wav.read(path_noisy_test)
dnn_magnitude_noisy_test, dnn_phase_noisy_test, _ = perform_stft(path_noisy_test, args)
dnn_magnitude_noisy_test_supervector = apply_context_window(dnn_magnitude_noisy_test.T, args.left_context, args.right_context).T
X_train = np.log(dnn_magnitude_noisy_train.T**2)
y_train = estimate_IRM(dnn_magnitude_clean_train.T,dnn_magnitude_noise_train.T)
X_test = np.log(dnn_magnitude_noisy_test_supervector.T**2)
# DNN training stage
#####################################################################################
k.clear_session()
def get_dnn_model(X_train, y_train, args):
# LeakyReLU, PReLU, ELU, ThresholdedReLU, SReLU
model = Sequential()
model.add(Dense(args.n_hidden, input_dim=X_train.shape[1], init='glorot_normal')) # glorot_normal,he_normal
model.add(BatchNormalization())
# model.add(Activation('relu'))
model.add(LeakyReLU(alpha=0.1))
model.add(Dropout(args.drop_out))
model.add(Dense(args.n_hidden, init='glorot_normal'))
model.add(BatchNormalization())
# model.add(Activation('relu'))
model.add(LeakyReLU(alpha=0.1))
model.add(Dropout(args.drop_out))
model.add(Dense(args.n_hidden, init='glorot_normal'))
model.add(BatchNormalization())
# model.add(Activation('relu'))
model.add(LeakyReLU(alpha=0.1))
model.add(Dropout(args.drop_out))
model.add(Dense(units=y_train.shape[1], init='glorot_normal'))
model.add(BatchNormalization())
model.add(Activation('linear'))
model.compile(loss='mse', optimizer='adam', metrics=['mse'])
# model.summary()
return model
model = get_dnn_model(X_train, y_train, args)
with tf.device('/gpu:0'):
model_info = model.fit(X_train, y_train, batch_size=args.n_batch, epochs=args.n_epoch)
# plot_model_history(model_info)
print("Training complete.")
# Enhancement stage
#####################################################################################
speech_masks = model.predict(X_test).T
magnitude_estimated_clean = dnn_magnitude_noisy_test * speech_masks
from numpy.linalg import norm
# magnitude_estimated_clean=norm(magnitude_estimated_clean)
# Reconstruction
stft_reconstructed_clean = merge_magphase(magnitude_estimated_clean, dnn_phase_noisy_test)
signal_reconstructed_clean =librosa.istft(stft_reconstructed_clean, hop_length=args.hop_size, window=args.window)
signal_reconstructed_clean = signal_reconstructed_clean.astype('int16')
#####################################################################################
output_path_estimated_noisy_test = os.path.join(PATH_ROOT, args.output_file)
wav.write(output_path_estimated_noisy_test,sr,signal_reconstructed_clean)
# Display signals, spectrograms
show_signal(clean_test,noisy_test,signal_reconstructed_clean,sr)
show_spectrogram(clean_test,noisy_test, signal_reconstructed_clean, sr, args.num_FFT,args.hop_size)
# # =============================================================================
# # PESQ
# # =============================================================================
# PATH_MATLAB='"C:/Program Files/MATLAB/R2014a/bin/matlab.exe"'
#
# PATH_MATLAB1 = os.path.join(PATH_ROOT , 'PESQ_MATLAB/execute_pesq.m')
# from pymatbridge import Matlab
# mlab = Matlab()
# mlab = Matlab(executable=PATH_MATLAB)
# mlab.start()
#
# #PATH_MATLAB1 = os.path.join(PATH_ROOT , "PESQ_MATLAB","execute_pesq.m")
# result_PESQ = mlab.run_func(PATH_MATLAB1, {'arg1': sr})
#
# noisy_original_PESQ = result_PESQ['result'][0][0]
# enhanced_PESQ = result_PESQ['result'][1][0]
#
# mlab.stop()
#
# snr=args.input_noisy_test
# name=snr[53:-9]
# print("[%s]\n Original: %.2f\n TF-Masking\t: %.2f"%(name,noisy_original_PESQ,enhanced_PESQ))
#
def parse_args():
parser = argparse.ArgumentParser(description='DNN-TF-Masking Speech Enhancement')
parser.add_argument('--datasets_dir', type=str, default='datasets/')
parser.add_argument('--input_clean_train', type=str, default='datasets/timit_clean_selected_train_total.wav')
parser.add_argument('--input_noisy_train', type=str, default='datasets/timit_noisy_selected_train_total.wav')
parser.add_argument('--input_noise', type=str, default='datasets/timit_noise_selected_total.wav')
parser.add_argument('--input_clean_test', type=str, default='datasets/timit_clean_selected/timit_clean_selected_test.wav')
parser.add_argument('--input_noisy_test', type=str, default='datasets/timit_noisy_selected/test_match/timit_noisy_babble_snr10_test.wav')
parser.add_argument('--input_dnn_clean_train', type=str, default='datasets/timit_clean_train_total.wav')
parser.add_argument('--input_dnn_noisy_train', type=str, default='datasets/timit_noisy_train_total.wav')
parser.add_argument('--input_dnn_noise', type=str, default='datasets/timit_noise_total.wav')
parser.add_argument('--output_file', type=str, default='datasets/output/estimated_clean_DNN-TF-Masking.wav')
parser.add_argument('--left-context', type=int, dest="left_context", default=0, help="left context of inputs for neural networks")
parser.add_argument('--right-context', type=int, dest="right_context", default=0, help="right context of inputs for neural networks")
parser.add_argument('--num_FFT', type=int, default='512', help='')
parser.add_argument('--hop_size', type=int, default='128', help='')
parser.add_argument('--window', type=str, default='hamming',help='boxcar, triang, blackman, hamming, hann, bartlett, flattop, parzen, bohman, blackmanharris, nuttall, barthann, kaiser')
parser.add_argument('--n_epoch', type=int, default='30', help='number of DNN epoch')
parser.add_argument('--n_hidden', type=int, default='1024', help='hidden units of DNN')
parser.add_argument('--drop_out', type=float, default='1', help='dropout of DNN')
parser.add_argument('--n_batch', type=int, default='1024', help='mini batch size')
return check_args(parser.parse_args())
def check_args(args):
if not os.path.exists(args.datasets_dir):
os.makedirs(args.datasets_dir)
assert args.num_FFT >= 1, 'number of FFT size must be larger than or equal to one'
assert args.hop_size < args.num_FFT, 'hop size must be smaller than number of FFT size'
return args
if __name__ == '__main__':
args = parse_args()
DNN_TF_Masking(args)