-
Notifications
You must be signed in to change notification settings - Fork 1
/
audio.py
161 lines (108 loc) · 4.78 KB
/
audio.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
import librosa
import librosa.filters
import math
import numpy as np
import tensorflow as tf
from scipy import signal
from hparams import create_hparams
hparams = create_hparams()
def load_wav(path):
return librosa.core.load(path, sr=hparams.sample_rate)[0]
def save_wav(wav, path):
wav *= 32767 / max(0.01, np.max(np.abs(wav)))
librosa.output.write_wav(path, wav.astype(np.int16), hparams.sample_rate)
def preemphasis(x):
return signal.lfilter([1, -hparams.preemphasis], [1], x)
def inv_preemphasis(x):
return signal.lfilter([1], [1, -hparams.preemphasis], x)
def spectrogram(y):
D = _stft(preemphasis(y))
S = _amp_to_db(np.abs(D)) - hparams.ref_level_db
return _normalize(S)
def inv_spectrogram(spectrogram):
'''Converts spectrogram to waveform using librosa'''
S = _db_to_amp(_denormalize(spectrogram) + hparams.ref_level_db) # Convert back to linear
return inv_preemphasis(_griffin_lim(S ** hparams.power)) # Reconstruct phase
def inv_spectrogram_tensorflow(spectrogram):
'''Builds computational graph to convert spectrogram to waveform using TensorFlow.
Unlike inv_spectrogram, this does NOT invert the preemphasis. The caller should call
inv_preemphasis on the output after running the graph.
'''
S = _db_to_amp_tensorflow(_denormalize_tensorflow(spectrogram) + hparams.ref_level_db)
return _griffin_lim_tensorflow(tf.pow(S, hparams.power))
def melspectrogram(y):
D = _stft(preemphasis(y))
S = _amp_to_db(_linear_to_mel(np.abs(D))) - hparams.ref_level_db
return _normalize(S)
def find_endpoint(wav, threshold_db=-40, min_silence_sec=0.8):
window_length = int(hparams.sample_rate * min_silence_sec)
hop_length = int(window_length / 4)
threshold = _db_to_amp(threshold_db)
for x in range(hop_length, len(wav) - window_length, hop_length):
if np.max(wav[x:x + window_length]) < threshold:
return x + hop_length
return len(wav)
def _griffin_lim(S):
'''librosa implementation of Griffin-Lim
Based on https://github.com/librosa/librosa/issues/434
'''
angles = np.exp(2j * np.pi * np.random.rand(*S.shape))
S_complex = np.abs(S).astype(np.complex)
y = _istft(S_complex * angles)
for i in range(hparams.griffin_lim_iters):
angles = np.exp(1j * np.angle(_stft(y)))
y = _istft(S_complex * angles)
return y
def _griffin_lim_tensorflow(S):
'''TensorFlow implementation of Griffin-Lim
Based on https://github.com/Kyubyong/tensorflow-exercises/blob/master/Audio_Processing.ipynb
'''
with tf.variable_scope('griffinlim'):
# TensorFlow's stft and istft operate on a batch of spectrograms; create batch of size 1
S = tf.expand_dims(S, 0)
S_complex = tf.identity(tf.cast(S, dtype=tf.complex64))
y = _istft_tensorflow(S_complex)
for i in range(hparams.griffin_lim_iters):
est = _stft_tensorflow(y)
angles = est / tf.cast(tf.maximum(1e-8, tf.abs(est)), tf.complex64)
y = _istft_tensorflow(S_complex * angles)
return tf.squeeze(y, 0)
def _stft(y):
n_fft, hop_length, win_length = _stft_parameters()
return librosa.stft(y=y, n_fft=n_fft, hop_length=hop_length, win_length=win_length)
def _istft(y):
_, hop_length, win_length = _stft_parameters()
return librosa.istft(y, hop_length=hop_length, win_length=win_length)
def _stft_tensorflow(signals):
n_fft, hop_length, win_length = _stft_parameters()
return tf.contrib.signal.stft(signals, win_length, hop_length, n_fft, pad_end=False)
def _istft_tensorflow(stfts):
n_fft, hop_length, win_length = _stft_parameters()
return tf.contrib.signal.inverse_stft(stfts, win_length, hop_length, n_fft)
def _stft_parameters():
n_fft = (hparams.num_freq - 1) * 2
hop_length = int(hparams.frame_shift_ms / 1000 * hparams.sample_rate)
win_length = int(hparams.frame_length_ms / 1000 * hparams.sample_rate)
return n_fft, hop_length, win_length
# Conversions:
_mel_basis = None
def _linear_to_mel(spectrogram):
global _mel_basis
if _mel_basis is None:
_mel_basis = _build_mel_basis()
return np.dot(_mel_basis, spectrogram)
def _build_mel_basis():
n_fft = (hparams.num_freq - 1) * 2
return librosa.filters.mel(hparams.sample_rate, n_fft, n_mels=hparams.num_mels)
def _amp_to_db(x):
return 20 * np.log10(np.maximum(1e-5, x))
def _db_to_amp(x):
return np.power(10.0, x * 0.05)
def _db_to_amp_tensorflow(x):
return tf.pow(tf.ones(tf.shape(x)) * 10.0, x * 0.05)
def _normalize(S):
return np.clip((S - hparams.min_level_db) / -hparams.min_level_db, 0, 1)
def _denormalize(S):
return (np.clip(S, 0, 1) * -hparams.min_level_db) + hparams.min_level_db
def _denormalize_tensorflow(S):
return (tf.clip_by_value(S, 0, 1) * -hparams.min_level_db) + hparams.min_level_db