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
|
# Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
# SPDX-License-Identifier: MIT
import os
import numpy as np
import pytest
import collections
from context import mfcc
from context import wav2letter_mfcc
from context import audio_capture
# Elements relevant to MFCC filter bank & feature extraction
MFCC_TEST_PARAMS = collections.namedtuple('mfcc_test_params',
['algo_params', 'mfcc_constructor', 'audio_proc_constructor'])
def kws_test_params():
kws_algo_params = mfcc.MFCCParams(sampling_freq=16000, num_fbank_bins=40, mel_lo_freq=20, mel_hi_freq=4000,
num_mfcc_feats=10, frame_len=640, use_htk_method=True, n_fft=1024)
return MFCC_TEST_PARAMS(kws_algo_params, mfcc.MFCC, mfcc.AudioPreprocessor)
def asr_test_params():
asr_algo_params = mfcc.MFCCParams(sampling_freq=16000, num_fbank_bins=128, mel_lo_freq=0, mel_hi_freq=8000,
num_mfcc_feats=13, frame_len=512, use_htk_method=False, n_fft=512)
return MFCC_TEST_PARAMS(asr_algo_params, wav2letter_mfcc.Wav2LetterMFCC, wav2letter_mfcc.W2LAudioPreprocessor)
def kws_cap_params():
return audio_capture.AudioCaptureParams(dtype=np.float32, overlap=0, min_samples=16000, sampling_freq=16000,
mono=True)
def asr_cap_params():
return audio_capture.AudioCaptureParams(dtype=np.float32, overlap=31712, min_samples=47712,
sampling_freq=16000, mono=True)
@pytest.fixture()
def audio_data(test_data_folder, file, audio_cap_params):
audio_file = os.path.join(test_data_folder, file)
capture = audio_capture.capture_audio(audio_file, audio_cap_params)
yield next(capture)
@pytest.mark.parametrize("file", ["yes.wav", "myVoiceIsMyPassportVerifyMe04.wav"])
@pytest.mark.parametrize("audio_cap_params", [kws_cap_params(), asr_cap_params()])
def test_audio_file(audio_data, test_data_folder, file, audio_cap_params):
assert audio_data.shape == (audio_cap_params.min_samples,)
assert audio_data.dtype == audio_cap_params.dtype
@pytest.mark.parametrize("mfcc_test_params, test_out", [(kws_test_params(), 25.470010570730597),
(asr_test_params(), 0.24)])
def test_mel_scale_function(mfcc_test_params, test_out):
mfcc_inst = mfcc_test_params.mfcc_constructor(mfcc_test_params.algo_params)
mel = mfcc_inst.mel_scale(16, mfcc_test_params.algo_params.use_htk_method)
assert np.isclose(mel, test_out)
@pytest.mark.parametrize("mfcc_test_params, test_out", [(kws_test_params(), 10.008767240008943),
(asr_test_params(), 1071.170287494467)])
def test_inverse_mel_scale_function(mfcc_test_params, test_out):
mfcc_inst = mfcc_test_params.mfcc_constructor(mfcc_test_params.algo_params)
mel = mfcc_inst.inv_mel_scale(16, mfcc_test_params.algo_params.use_htk_method)
assert np.isclose(mel, test_out)
mel_filter_test_data_kws = {0: [0.33883214, 0.80088392, 0.74663128, 0.30332531],
1: [0.25336872, 0.69667469, 0.86883317, 0.44281119, 0.02493546],
2: [0.13116683, 0.55718881, 0.97506454, 0.61490026, 0.21241678],
5: [0.32725038, 0.69579596, 0.9417706, 0.58524989, 0.23445207],
-1: [0.02433275, 0.10371618, 0.1828123, 0.26162319, 0.34015089, 0.41839743,
0.49636481, 0.57405503, 0.65147004, 0.72861179, 0.8054822, 0.88208318,
0.95841659, 0.96551568, 0.88971181, 0.81416996, 0.73888833, 0.66386514,
0.58909861, 0.514587, 0.44032856, 0.3663216, 0.29256441, 0.21905531,
0.14579264, 0.07277474]}
mel_filter_test_data_asr = {0: [0.02837754],
1: [0.01438901, 0.01398853],
2: [0.02877802],
5: [0.01478948, 0.01358806],
-1: [4.82151203e-05, 9.48791110e-04, 1.84569875e-03, 2.73896782e-03,
3.62862771e-03, 4.51470746e-03, 5.22215439e-03, 4.34314914e-03,
3.46763895e-03, 2.59559614e-03, 1.72699334e-03, 8.61803536e-04]}
@pytest.mark.parametrize("mfcc_test_params, test_out",
[(kws_test_params(), mel_filter_test_data_kws),
(asr_test_params(), mel_filter_test_data_asr)])
def test_create_mel_filter_bank(mfcc_test_params, test_out):
mfcc_inst = mfcc_test_params.mfcc_constructor(mfcc_test_params.algo_params)
mel_filter_bank = mfcc_inst.create_mel_filter_bank()
assert len(mel_filter_bank) == mfcc_test_params.algo_params.num_fbank_bins
for indx, data in test_out.items():
assert np.allclose(mel_filter_bank[indx], data)
mfcc_test_data_kws = (-22.671347398982626, -0.6161543999707211, 2.072326974167832,
0.5813741475362223, 1.0165529747334272, 0.8581560719988703,
0.4603911069624896, 0.03392820944377398, 1.1651093266902361,
0.007200025869960908)
mfcc_test_data_asr = (-735.46345398, 69.50331943, 16.39159347, 22.74874819, 24.84782893,
10.67559303, 12.82828618, -3.51084271, 4.66633677, 10.20079095, 11.34782948, 3.90499354,
9.32322384)
@pytest.mark.parametrize("mfcc_test_params, test_out, file, audio_cap_params",
[(kws_test_params(), mfcc_test_data_kws, "yes.wav", kws_cap_params()),
(asr_test_params(), mfcc_test_data_asr, "myVoiceIsMyPassportVerifyMe04.wav",
asr_cap_params())])
def test_mfcc_compute_first_frame(audio_data, mfcc_test_params, test_out, file, audio_cap_params):
audio_data = np.array(audio_data)[0:mfcc_test_params.algo_params.frame_len]
mfcc_inst = mfcc_test_params.mfcc_constructor(mfcc_test_params.algo_params)
mfcc_feats = mfcc_inst.mfcc_compute(audio_data)
assert np.allclose((mfcc_feats[0:mfcc_test_params.algo_params.num_mfcc_feats]), test_out)
extract_test_data_kws = {0: [-2.2671347e+01, -6.1615437e-01, 2.0723269e+00, 5.8137417e-01,
1.0165529e+00, 8.5815609e-01, 4.6039110e-01, 3.3928208e-02,
1.1651093e+00, 7.2000260e-03],
1: [-23.488806, -1.1687667, 3.0548365, 1.5129884, 1.4142203,
0.6869772, 1.1875846, 0.5743369, 1.202258, -0.12133602],
2: [-23.909292, -1.5186096, 1.8721082, 0.7378916, 0.44974303,
0.17609395, 0.5183161, 0.37109664, 0.14186797, 0.58400506],
-1: [-23.752186, -0.1796912, 1.9514247, 0.32554424, 1.8425112,
0.8763608, 0.78326845, 0.27808753, 0.73788685, 0.30338883]}
extract_test_data_asr = {0: [-4.98830318e+00, 6.86444461e-01, 3.12024504e-01, 3.56840312e-01,
3.71638149e-01, 2.71728605e-01, 2.86904365e-01, 1.71718955e-01,
2.29365349e-01, 2.68381387e-01, 2.76467651e-01, 2.23998129e-01,
2.62194842e-01, -1.48247385e+01, 1.21875501e+00, 4.20235842e-01,
5.39400637e-01, 6.09882712e-01, 1.68513224e-01, 3.75330061e-01,
8.57576132e-02, 1.92831963e-01, 1.41814977e-01, 1.57615796e-01,
7.19076321e-02, 1.98729336e-01, 3.92199278e+00, -5.76856315e-01,
1.17938723e-02, -9.25096497e-02, -3.59488949e-02, 1.13284402e-03,
1.51282102e-01, 1.13404110e-01, -8.69824737e-02, -1.48449212e-01,
-1.24230251e-01, -1.90728232e-01, -5.37525006e-02],
1: [-4.96694946e+00, 6.69411421e-01, 2.86189795e-01, 3.65071595e-01,
3.92671198e-01, 2.44258150e-01, 2.52177566e-01, 2.16024980e-01,
2.79812217e-01, 2.79687315e-01, 2.95228422e-01, 2.83991724e-01,
2.46358261e-01, -1.33618221e+01, 1.08920455e+00, 3.88707787e-01,
5.05674303e-01, 6.08285785e-01, 1.68113053e-01, 3.54529470e-01,
6.68609440e-02, 1.52882755e-01, 6.89579248e-02, 1.18375972e-01,
5.86742274e-02, 1.15678251e-01, 1.07892036e+01, -1.07193100e+00,
-2.18140319e-01, -3.35950345e-01, -2.57241666e-01, -5.54431602e-02,
-8.38544443e-02, -5.79114584e-03, -2.23973781e-01, -2.91451365e-01,
-2.11069033e-01, -1.90297231e-01, -2.76504964e-01],
2: [-4.98664522e+00, 6.54802263e-01, 3.70355755e-01, 4.06837821e-01,
4.05175537e-01, 2.29149669e-01, 2.83312678e-01, 2.17573136e-01,
3.07824671e-01, 2.48388007e-01, 2.25399241e-01, 2.52003014e-01,
2.83968121e-01, -1.05043650e+01, 7.91533887e-01, 3.11546475e-01,
4.36079264e-01, 5.93271911e-01, 2.02480286e-01, 3.24254721e-01,
6.29674867e-02, 9.67641100e-02, -1.62826646e-02, 5.47595806e-02,
2.90475693e-02, 2.62522381e-02, 1.38787737e+01, -1.32597208e+00,
-3.73900205e-01, -4.38065380e-01, -3.05983245e-01, 1.14390980e-02,
-2.10821658e-01, -6.22789040e-02, -2.88273603e-01, -3.29794526e-01,
-2.43764088e-01, -1.70954674e-01, -3.65193188e-01],
-1: [-2.1894817, 1.583355, -0.45024827, 0.11657667, 0.08940444, 0.09041209,
0.2003613, 0.11800499, 0.18838657, 0.29271516, 0.22758003, 0.10634928,
-0.04019014, 7.203311, -2.414309, 0.28750962, -0.24222863, 0.04680864,
-0.12129474, 0.18059334, 0.06250379, 0.11363743, -0.2561094, -0.08132717,
-0.08500769, 0.18916495, 1.3529671, -3.7919693, 1.937804, 0.6845761,
0.15381853, 0.41106734, -0.28207013, 0.2195526, 0.06716935, -0.02886542,
-0.22860551, 0.24788341, 0.63940096]}
@pytest.mark.parametrize("mfcc_test_params, model_input_size, stride, min_samples, file, audio_cap_params, test_out",
[(kws_test_params(), 49, 320, 16000, "yes.wav", kws_cap_params(),
extract_test_data_kws),
(asr_test_params(), 296, 160, 47712, "myVoiceIsMyPassportVerifyMe04.wav", asr_cap_params(),
extract_test_data_asr)])
def test_feat_extraction_full_sized_input(audio_data,
mfcc_test_params,
model_input_size,
stride,
min_samples, file, audio_cap_params,
test_out):
"""
Test out values were gathered by printing the mfcc features collected during the first full inference
on the test wav files. Note the extract_features() function simply calls the mfcc_compute() from previous
test but feeds in enough samples for an inference rather than a single frame. It also computes the 1st & 2nd
derivative features hence the shape (13*3 = 39).
Specific model_input_size and stride parameters are also required as additional arguments.
"""
audio_data = np.array(audio_data)
# Pad with zeros to ensure min_samples for inference
audio_data.resize(min_samples)
mfcc_inst = mfcc_test_params.mfcc_constructor(mfcc_test_params.algo_params)
preprocessor = mfcc_test_params.audio_proc_constructor(mfcc_inst, model_input_size, stride)
# extract_features passes the audio data to mfcc_compute frame by frame and concatenates results
input_tensor = preprocessor.extract_features(audio_data)
assert len(input_tensor) == model_input_size
for indx, data in test_out.items():
assert np.allclose(input_tensor[indx], data)
# Expected contents of input tensors for inference on a silent wav file
extract_features_zeros_kws = {0: [-2.05949466e+02, -4.88498131e-15, 8.15428020e-15, -5.77315973e-15,
7.03142511e-15, -1.11022302e-14, 2.18015108e-14, -1.77635684e-15,
1.06581410e-14, 2.75335310e-14],
-1: [-2.05949466e+02, -4.88498131e-15, 8.15428020e-15, -5.77315973e-15,
7.03142511e-15, -1.11022302e-14, 2.18015108e-14, -1.77635684e-15,
1.06581410e-14, 2.75335310e-14]}
extract_features_zeros_asr = {
0: [-3.46410162e+00, 2.88675135e-01, 2.88675135e-01, 2.88675135e-01,
2.88675135e-01, 2.88675135e-01, 2.88675135e-01, 2.88675135e-01,
2.88675135e-01, 2.88675135e-01, 2.88675135e-01, 2.88675135e-01,
2.88675135e-01, 2.79662980e+01, 1.75638694e-15, -9.41313626e-16,
9.66012817e-16, -1.23221521e-15, 1.75638694e-15, -1.59035349e-15,
2.41503204e-15, -1.64798493e-15, 4.39096735e-16, -4.95356004e-16,
-2.19548368e-16, -3.55668355e-15, 8.19843971e+00, -4.28340672e-02,
-4.28340672e-02, -4.28340672e-02, -4.28340672e-02, -4.28340672e-02,
-4.28340672e-02, -4.28340672e-02, -4.28340672e-02, -4.28340672e-02,
-4.28340672e-02, -4.28340672e-02, -4.28340672e-02],
- 1: [-3.46410162e+00, 2.88675135e-01, 2.88675135e-01, 2.88675135e-01,
2.88675135e-01, 2.88675135e-01, 2.88675135e-01, 2.88675135e-01,
2.88675135e-01, 2.88675135e-01, 2.88675135e-01, 2.88675135e-01,
2.88675135e-01, 2.79662980e+01, 1.75638694e-15, -9.41313626e-16,
9.66012817e-16, -1.23221521e-15, 1.75638694e-15, -1.59035349e-15,
2.41503204e-15, -1.64798493e-15, 4.39096735e-16, -4.95356004e-16,
-2.19548368e-16, -3.55668355e-15, 8.19843971e+00, -4.28340672e-02,
-4.28340672e-02, -4.28340672e-02, -4.28340672e-02, -4.28340672e-02,
-4.28340672e-02, -4.28340672e-02, -4.28340672e-02, -4.28340672e-02,
-4.28340672e-02, -4.28340672e-02, -4.28340672e-02]}
@pytest.mark.parametrize("mfcc_test_params,model_input_size, stride, min_samples, test_out",
[(kws_test_params(), 49, 320, 16000, extract_features_zeros_kws),
(asr_test_params(), 296, 160, 47712, extract_features_zeros_asr)])
def test_feat_extraction_full_sized_input_zeros(mfcc_test_params, model_input_size, stride, min_samples, test_out):
audio_data = np.zeros(min_samples).astype(np.float32)
mfcc_inst = mfcc_test_params.mfcc_constructor(mfcc_test_params.algo_params)
preprocessor = mfcc_test_params.audio_proc_constructor(mfcc_inst, model_input_size,
stride)
input_tensor = preprocessor.extract_features(audio_data)
assert len(input_tensor) == model_input_size
for indx, data in test_out.items():
# Element 14 of feature extraction vector differs minutely during
# inference on a silent wav file compared to array of 0's
# Workarounds were to skip this sample or add large tolerance argument (atol=10)
assert np.allclose(input_tensor[indx][0:13], data[0:13])
assert np.allclose(input_tensor[indx][15:], data[15:])
|
