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
|
# -*- coding: utf-8 -*-
"""
Audio Feature Augmentation
==========================
**Author**: `Moto Hira <moto@meta.com>`__
"""
# When running this tutorial in Google Colab, install the required packages
# with the following.
# !pip install torchaudio librosa
import torch
import torchaudio
import torchaudio.transforms as T
print(torch.__version__)
print(torchaudio.__version__)
######################################################################
# Preparation
# -----------
#
import librosa
import matplotlib.pyplot as plt
from IPython.display import Audio
from torchaudio.utils import download_asset
######################################################################
# In this tutorial, we will use a speech data from
# `VOiCES dataset <https://iqtlabs.github.io/voices/>`__,
# which is licensed under Creative Commos BY 4.0.
SAMPLE_WAV_SPEECH_PATH = download_asset("tutorial-assets/Lab41-SRI-VOiCES-src-sp0307-ch127535-sg0042.wav")
def _get_sample(path, resample=None):
effects = [["remix", "1"]]
if resample:
effects.extend(
[
["lowpass", f"{resample // 2}"],
["rate", f"{resample}"],
]
)
return torchaudio.sox_effects.apply_effects_file(path, effects=effects)
def get_speech_sample(*, resample=None):
return _get_sample(SAMPLE_WAV_SPEECH_PATH, resample=resample)
def get_spectrogram(
n_fft=400,
win_len=None,
hop_len=None,
power=2.0,
):
waveform, _ = get_speech_sample()
spectrogram = T.Spectrogram(
n_fft=n_fft,
win_length=win_len,
hop_length=hop_len,
center=True,
pad_mode="reflect",
power=power,
)
return spectrogram(waveform)
######################################################################
# SpecAugment
# -----------
#
# `SpecAugment <https://ai.googleblog.com/2019/04/specaugment-new-data-augmentation.html>`__
# is a popular spectrogram augmentation technique.
#
# ``torchaudio`` implements :py:func:`torchaudio.transforms.TimeStretch`,
# :py:func:`torchaudio.transforms.TimeMasking` and
# :py:func:`torchaudio.transforms.FrequencyMasking`.
#
######################################################################
# TimeStretch
# -----------
#
spec = get_spectrogram(power=None)
stretch = T.TimeStretch()
spec_12 = stretch(spec, overriding_rate=1.2)
spec_09 = stretch(spec, overriding_rate=0.9)
######################################################################
# Visualization
# ~~~~~~~~~~~~~
def plot():
def plot_spec(ax, spec, title):
ax.set_title(title)
ax.imshow(librosa.amplitude_to_db(spec), origin="lower", aspect="auto")
fig, axes = plt.subplots(3, 1, sharex=True, sharey=True)
plot_spec(axes[0], torch.abs(spec_12[0]), title="Stretched x1.2")
plot_spec(axes[1], torch.abs(spec[0]), title="Original")
plot_spec(axes[2], torch.abs(spec_09[0]), title="Stretched x0.9")
fig.tight_layout()
plot()
######################################################################
# Audio Samples
# ~~~~~~~~~~~~~
def preview(spec, rate=16000):
ispec = T.InverseSpectrogram()
waveform = ispec(spec)
return Audio(waveform[0].numpy().T, rate=rate)
preview(spec)
######################################################################
#
preview(spec_12)
######################################################################
#
preview(spec_09)
######################################################################
# Time and Frequency Masking
# --------------------------
#
torch.random.manual_seed(4)
time_masking = T.TimeMasking(time_mask_param=80)
freq_masking = T.FrequencyMasking(freq_mask_param=80)
spec = get_spectrogram()
time_masked = time_masking(spec)
freq_masked = freq_masking(spec)
######################################################################
#
def plot():
def plot_spec(ax, spec, title):
ax.set_title(title)
ax.imshow(librosa.power_to_db(spec), origin="lower", aspect="auto")
fig, axes = plt.subplots(3, 1, sharex=True, sharey=True)
plot_spec(axes[0], spec[0], title="Original")
plot_spec(axes[1], time_masked[0], title="Masked along time axis")
plot_spec(axes[2], freq_masked[0], title="Masked along frequency axis")
fig.tight_layout()
plot()
|
