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
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
|
import itertools
from torchaudio.backend import sox_io_backend
from parameterized import parameterized
from torchaudio_unittest.common_utils import (
TempDirMixin,
PytorchTestCase,
skipIfNoExec,
skipIfNoExtension,
get_wav_data,
load_wav,
save_wav,
sox_utils,
)
from .common import (
name_func,
)
class SaveTestBase(TempDirMixin, PytorchTestCase):
def assert_wav(self, dtype, sample_rate, num_channels, num_frames):
"""`sox_io_backend.save` can save wav format."""
path = self.get_temp_path('data.wav')
expected = get_wav_data(dtype, num_channels, num_frames=num_frames)
sox_io_backend.save(path, expected, sample_rate)
found, sr = load_wav(path)
assert sample_rate == sr
self.assertEqual(found, expected)
def assert_mp3(self, sample_rate, num_channels, bit_rate, duration):
"""`sox_io_backend.save` can save mp3 format.
mp3 encoding introduces delay and boundary effects so
we convert the resulting mp3 to wav and compare the results there
|
| 1. Generate original wav file with SciPy
|
v
-------------- wav ----------------
| |
| 2.1. load with scipy | 3.1. Convert to mp3 with Sox
| then save with torchaudio |
v v
mp3 mp3
| |
| 2.2. Convert to wav with Sox | 3.2. Convert to wav with Sox
| |
v v
wav wav
| |
| 2.3. load with scipy | 3.3. load with scipy
| |
v v
tensor -------> compare <--------- tensor
"""
src_path = self.get_temp_path('1.reference.wav')
mp3_path = self.get_temp_path('2.1.torchaudio.mp3')
wav_path = self.get_temp_path('2.2.torchaudio.wav')
mp3_path_sox = self.get_temp_path('3.1.sox.mp3')
wav_path_sox = self.get_temp_path('3.2.sox.wav')
# 1. Generate original wav
data = get_wav_data('float32', num_channels, normalize=True, num_frames=duration * sample_rate)
save_wav(src_path, data, sample_rate)
# 2.1. Convert the original wav to mp3 with torchaudio
sox_io_backend.save(
mp3_path, load_wav(src_path)[0], sample_rate, compression=bit_rate)
# 2.2. Convert the mp3 to wav with Sox
sox_utils.convert_audio_file(mp3_path, wav_path)
# 2.3. Load
found = load_wav(wav_path)[0]
# 3.1. Convert the original wav to mp3 with SoX
sox_utils.convert_audio_file(src_path, mp3_path_sox, compression=bit_rate)
# 3.2. Convert the mp3 to wav with Sox
sox_utils.convert_audio_file(mp3_path_sox, wav_path_sox)
# 3.3. Load
expected = load_wav(wav_path_sox)[0]
self.assertEqual(found, expected)
def assert_flac(self, sample_rate, num_channels, compression_level, duration):
"""`sox_io_backend.save` can save flac format.
This test takes the same strategy as mp3 to compare the result
"""
src_path = self.get_temp_path('1.reference.wav')
flc_path = self.get_temp_path('2.1.torchaudio.flac')
wav_path = self.get_temp_path('2.2.torchaudio.wav')
flc_path_sox = self.get_temp_path('3.1.sox.flac')
wav_path_sox = self.get_temp_path('3.2.sox.wav')
# 1. Generate original wav
data = get_wav_data('float32', num_channels, normalize=True, num_frames=duration * sample_rate)
save_wav(src_path, data, sample_rate)
# 2.1. Convert the original wav to flac with torchaudio
sox_io_backend.save(
flc_path, load_wav(src_path)[0], sample_rate, compression=compression_level)
# 2.2. Convert the flac to wav with Sox
# converting to 32 bit because flac file has 24 bit depth which scipy cannot handle.
sox_utils.convert_audio_file(flc_path, wav_path, bit_depth=32)
# 2.3. Load
found = load_wav(wav_path)[0]
# 3.1. Convert the original wav to flac with SoX
sox_utils.convert_audio_file(src_path, flc_path_sox, compression=compression_level)
# 3.2. Convert the flac to wav with Sox
# converting to 32 bit because flac file has 24 bit depth which scipy cannot handle.
sox_utils.convert_audio_file(flc_path_sox, wav_path_sox, bit_depth=32)
# 3.3. Load
expected = load_wav(wav_path_sox)[0]
self.assertEqual(found, expected)
def _assert_vorbis(self, sample_rate, num_channels, quality_level, duration):
"""`sox_io_backend.save` can save vorbis format.
This test takes the same strategy as mp3 to compare the result
"""
src_path = self.get_temp_path('1.reference.wav')
vbs_path = self.get_temp_path('2.1.torchaudio.vorbis')
wav_path = self.get_temp_path('2.2.torchaudio.wav')
vbs_path_sox = self.get_temp_path('3.1.sox.vorbis')
wav_path_sox = self.get_temp_path('3.2.sox.wav')
# 1. Generate original wav
data = get_wav_data('int16', num_channels, normalize=False, num_frames=duration * sample_rate)
save_wav(src_path, data, sample_rate)
# 2.1. Convert the original wav to vorbis with torchaudio
sox_io_backend.save(
vbs_path, load_wav(src_path)[0], sample_rate, compression=quality_level)
# 2.2. Convert the vorbis to wav with Sox
sox_utils.convert_audio_file(vbs_path, wav_path)
# 2.3. Load
found = load_wav(wav_path)[0]
# 3.1. Convert the original wav to vorbis with SoX
sox_utils.convert_audio_file(src_path, vbs_path_sox, compression=quality_level)
# 3.2. Convert the vorbis to wav with Sox
sox_utils.convert_audio_file(vbs_path_sox, wav_path_sox)
# 3.3. Load
expected = load_wav(wav_path_sox)[0]
# sox's vorbis encoding has some random boundary effect, which cause small number of
# samples yields higher descrepency than the others.
# so we allow small portions of data to be outside of absolute torelance.
# make sure to pass somewhat long duration
atol = 1.0e-4
max_failure_allowed = 0.01 # this percent of samples are allowed to outside of atol.
failure_ratio = ((found - expected).abs() > atol).sum().item() / found.numel()
if failure_ratio > max_failure_allowed:
# it's failed and this will give a better error message.
self.assertEqual(found, expected, atol=atol, rtol=1.3e-6)
def assert_vorbis(self, *args, **kwargs):
# sox's vorbis encoding has some randomness, so we run tests multiple time
max_retry = 5
error = None
for _ in range(max_retry):
try:
self._assert_vorbis(*args, **kwargs)
break
except AssertionError as e:
error = e
else:
raise error
def assert_sphere(self, sample_rate, num_channels, duration):
"""`sox_io_backend.save` can save sph format.
This test takes the same strategy as mp3 to compare the result
"""
src_path = self.get_temp_path('1.reference.wav')
flc_path = self.get_temp_path('2.1.torchaudio.sph')
wav_path = self.get_temp_path('2.2.torchaudio.wav')
flc_path_sox = self.get_temp_path('3.1.sox.sph')
wav_path_sox = self.get_temp_path('3.2.sox.wav')
# 1. Generate original wav
data = get_wav_data('float32', num_channels, normalize=True, num_frames=duration * sample_rate)
save_wav(src_path, data, sample_rate)
# 2.1. Convert the original wav to sph with torchaudio
sox_io_backend.save(flc_path, load_wav(src_path)[0], sample_rate)
# 2.2. Convert the sph to wav with Sox
# converting to 32 bit because sph file has 24 bit depth which scipy cannot handle.
sox_utils.convert_audio_file(flc_path, wav_path, bit_depth=32)
# 2.3. Load
found = load_wav(wav_path)[0]
# 3.1. Convert the original wav to sph with SoX
sox_utils.convert_audio_file(src_path, flc_path_sox)
# 3.2. Convert the sph to wav with Sox
# converting to 32 bit because sph file has 24 bit depth which scipy cannot handle.
sox_utils.convert_audio_file(flc_path_sox, wav_path_sox, bit_depth=32)
# 3.3. Load
expected = load_wav(wav_path_sox)[0]
self.assertEqual(found, expected)
@skipIfNoExec('sox')
@skipIfNoExtension
class TestSave(SaveTestBase):
@parameterized.expand(list(itertools.product(
['float32', 'int32', 'int16', 'uint8'],
[8000, 16000],
[1, 2],
)), name_func=name_func)
def test_wav(self, dtype, sample_rate, num_channels):
"""`sox_io_backend.save` can save wav format."""
self.assert_wav(dtype, sample_rate, num_channels, num_frames=None)
@parameterized.expand(list(itertools.product(
['float32'],
[16000],
[2],
)), name_func=name_func)
def test_wav_large(self, dtype, sample_rate, num_channels):
"""`sox_io_backend.save` can save large wav file."""
two_hours = 2 * 60 * 60 * sample_rate
self.assert_wav(dtype, sample_rate, num_channels, num_frames=two_hours)
@parameterized.expand(list(itertools.product(
['float32', 'int32', 'int16', 'uint8'],
[4, 8, 16, 32],
)), name_func=name_func)
def test_multiple_channels(self, dtype, num_channels):
"""`sox_io_backend.save` can save wav with more than 2 channels."""
sample_rate = 8000
self.assert_wav(dtype, sample_rate, num_channels, num_frames=None)
@parameterized.expand(list(itertools.product(
[8000, 16000],
[1, 2],
[-4.2, -0.2, 0, 0.2, 96, 128, 160, 192, 224, 256, 320],
)), name_func=name_func)
def test_mp3(self, sample_rate, num_channels, bit_rate):
"""`sox_io_backend.save` can save mp3 format."""
self.assert_mp3(sample_rate, num_channels, bit_rate, duration=1)
@parameterized.expand(list(itertools.product(
[16000],
[2],
[128],
)), name_func=name_func)
def test_mp3_large(self, sample_rate, num_channels, bit_rate):
"""`sox_io_backend.save` can save large mp3 file."""
two_hours = 2 * 60 * 60
self.assert_mp3(sample_rate, num_channels, bit_rate, duration=two_hours)
@parameterized.expand(list(itertools.product(
[8000, 16000],
[1, 2],
[None] + list(range(9)),
)), name_func=name_func)
def test_flac(self, sample_rate, num_channels, compression_level):
"""`sox_io_backend.save` can save flac format."""
self.assert_flac(sample_rate, num_channels, compression_level, duration=1)
@parameterized.expand(list(itertools.product(
[16000],
[2],
[0],
)), name_func=name_func)
def test_flac_large(self, sample_rate, num_channels, compression_level):
"""`sox_io_backend.save` can save large flac file."""
two_hours = 2 * 60 * 60
self.assert_flac(sample_rate, num_channels, compression_level, duration=two_hours)
@parameterized.expand(list(itertools.product(
[8000, 16000],
[1, 2],
[None, -1, 0, 1, 2, 3, 3.6, 5, 10],
)), name_func=name_func)
def test_vorbis(self, sample_rate, num_channels, quality_level):
"""`sox_io_backend.save` can save vorbis format."""
self.assert_vorbis(sample_rate, num_channels, quality_level, duration=20)
# note: torchaudio can load large vorbis file, but cannot save large volbis file
# the following test causes Segmentation fault
#
'''
@parameterized.expand(list(itertools.product(
[16000],
[2],
[10],
)), name_func=name_func)
def test_vorbis_large(self, sample_rate, num_channels, quality_level):
"""`sox_io_backend.save` can save large vorbis file correctly."""
two_hours = 2 * 60 * 60
self.assert_vorbis(sample_rate, num_channels, quality_level, two_hours)
'''
@parameterized.expand(list(itertools.product(
[8000, 16000],
[1, 2],
)), name_func=name_func)
def test_sphere(self, sample_rate, num_channels):
"""`sox_io_backend.save` can save sph format."""
self.assert_sphere(sample_rate, num_channels, duration=1)
@skipIfNoExec('sox')
@skipIfNoExtension
class TestSaveParams(TempDirMixin, PytorchTestCase):
"""Test the correctness of optional parameters of `sox_io_backend.save`"""
@parameterized.expand([(True, ), (False, )], name_func=name_func)
def test_channels_first(self, channels_first):
"""channels_first swaps axes"""
path = self.get_temp_path('data.wav')
data = get_wav_data('int32', 2, channels_first=channels_first)
sox_io_backend.save(
path, data, 8000, channels_first=channels_first)
found = load_wav(path)[0]
expected = data if channels_first else data.transpose(1, 0)
self.assertEqual(found, expected)
@parameterized.expand([
'float32', 'int32', 'int16', 'uint8'
], name_func=name_func)
def test_noncontiguous(self, dtype):
"""Noncontiguous tensors are saved correctly"""
path = self.get_temp_path('data.wav')
expected = get_wav_data(dtype, 4)[::2, ::2]
assert not expected.is_contiguous()
sox_io_backend.save(path, expected, 8000)
found = load_wav(path)[0]
self.assertEqual(found, expected)
@parameterized.expand([
'float32', 'int32', 'int16', 'uint8',
])
def test_tensor_preserve(self, dtype):
"""save function should not alter Tensor"""
path = self.get_temp_path('data.wav')
expected = get_wav_data(dtype, 4)[::2, ::2]
data = expected.clone()
sox_io_backend.save(path, data, 8000)
self.assertEqual(data, expected)
|
