-
Notifications
You must be signed in to change notification settings - Fork 126
/
data_tools.py
133 lines (91 loc) · 4.96 KB
/
data_tools.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
import librosa
import numpy as np
import os
def audio_to_audio_frame_stack(sound_data, frame_length, hop_length_frame):
"""This function take an audio and split into several frame
in a numpy matrix of size (nb_frame,frame_length)"""
sequence_sample_length = sound_data.shape[0]
sound_data_list = [sound_data[start:start + frame_length] for start in range(
0, sequence_sample_length - frame_length + 1, hop_length_frame)] # get sliding windows
sound_data_array = np.vstack(sound_data_list)
return sound_data_array
def audio_files_to_numpy(audio_dir, list_audio_files, sample_rate, frame_length, hop_length_frame, min_duration):
"""This function take audio files of a directory and merge them
in a numpy matrix of size (nb_frame,frame_length) for a sliding window of size hop_length_frame"""
list_sound_array = []
for file in list_audio_files:
# open the audio file
y, sr = librosa.load(os.path.join(audio_dir, file), sr=sample_rate)
total_duration = librosa.get_duration(y=y, sr=sr)
if (total_duration >= min_duration):
list_sound_array.append(audio_to_audio_frame_stack(
y, frame_length, hop_length_frame))
else:
print(
f"The following file {os.path.join(audio_dir,file)} is below the min duration")
return np.vstack(list_sound_array)
def blend_noise_randomly(voice, noise, nb_samples, frame_length):
"""This function takes as input numpy arrays representing frames
of voice sounds, noise sounds and the number of frames to be created
and return numpy arrays with voice randomly blend with noise"""
prod_voice = np.zeros((nb_samples, frame_length))
prod_noise = np.zeros((nb_samples, frame_length))
prod_noisy_voice = np.zeros((nb_samples, frame_length))
for i in range(nb_samples):
id_voice = np.random.randint(0, voice.shape[0])
id_noise = np.random.randint(0, noise.shape[0])
level_noise = np.random.uniform(0.2, 0.8)
prod_voice[i, :] = voice[id_voice, :]
prod_noise[i, :] = level_noise * noise[id_noise, :]
prod_noisy_voice[i, :] = prod_voice[i, :] + prod_noise[i, :]
return prod_voice, prod_noise, prod_noisy_voice
def audio_to_magnitude_db_and_phase(n_fft, hop_length_fft, audio):
"""This function takes an audio and convert into spectrogram,
it returns the magnitude in dB and the phase"""
stftaudio = librosa.stft(audio, n_fft=n_fft, hop_length=hop_length_fft)
stftaudio_magnitude, stftaudio_phase = librosa.magphase(stftaudio)
stftaudio_magnitude_db = librosa.amplitude_to_db(
stftaudio_magnitude, ref=np.max)
return stftaudio_magnitude_db, stftaudio_phase
def numpy_audio_to_matrix_spectrogram(numpy_audio, dim_square_spec, n_fft, hop_length_fft):
"""This function takes as input a numpy audi of size (nb_frame,frame_length), and return
a numpy containing the matrix spectrogram for amplitude in dB and phase. It will have the size
(nb_frame,dim_square_spec,dim_square_spec)"""
nb_audio = numpy_audio.shape[0]
m_mag_db = np.zeros((nb_audio, dim_square_spec, dim_square_spec))
m_phase = np.zeros((nb_audio, dim_square_spec, dim_square_spec), dtype=complex)
for i in range(nb_audio):
m_mag_db[i, :, :], m_phase[i, :, :] = audio_to_magnitude_db_and_phase(
n_fft, hop_length_fft, numpy_audio[i])
return m_mag_db, m_phase
def magnitude_db_and_phase_to_audio(frame_length, hop_length_fft, stftaudio_magnitude_db, stftaudio_phase):
"""This functions reverts a spectrogram to an audio"""
stftaudio_magnitude_rev = librosa.db_to_amplitude(stftaudio_magnitude_db, ref=1.0)
# taking magnitude and phase of audio
audio_reverse_stft = stftaudio_magnitude_rev * stftaudio_phase
audio_reconstruct = librosa.core.istft(audio_reverse_stft, hop_length=hop_length_fft, length=frame_length)
return audio_reconstruct
def matrix_spectrogram_to_numpy_audio(m_mag_db, m_phase, frame_length, hop_length_fft) :
"""This functions reverts the matrix spectrograms to numpy audio"""
list_audio = []
nb_spec = m_mag_db.shape[0]
for i in range(nb_spec):
audio_reconstruct = magnitude_db_and_phase_to_audio(frame_length, hop_length_fft, m_mag_db[i], m_phase[i])
list_audio.append(audio_reconstruct)
return np.vstack(list_audio)
def scaled_in(matrix_spec):
"global scaling apply to noisy voice spectrograms (scale between -1 and 1)"
matrix_spec = (matrix_spec + 46)/50
return matrix_spec
def scaled_ou(matrix_spec):
"global scaling apply to noise models spectrograms (scale between -1 and 1)"
matrix_spec = (matrix_spec -6 )/82
return matrix_spec
def inv_scaled_in(matrix_spec):
"inverse global scaling apply to noisy voices spectrograms"
matrix_spec = matrix_spec * 50 - 46
return matrix_spec
def inv_scaled_ou(matrix_spec):
"inverse global scaling apply to noise models spectrograms"
matrix_spec = matrix_spec * 82 + 6
return matrix_spec