def test_gain_pitch_same_size():
"""
tests that varying the gain and the pitch has no affect on the audio_data size
"""
tempo = 1.0
gain_pitch_tuples = [
(0, 0), # not augmentation
(8, 0), # only gain aug
(0, 400), # only pitch
(-6, -400)
] # both gain and pitch
audio_path = get_all_test_audio()[0] # only using a single audio path
for gain, pitch in gain_pitch_tuples:
# un-augmented audio_data
audio_data, samp_rate = array_from_wave(audio_path)
aug_data, samp_rate = tempo_gain_pitch_perturb(audio_path,
sample_rate=samp_rate,
tempo_range=(tempo,
tempo),
gain_range=(gain, gain),
pitch_range=(pitch,
pitch))
assert audio_data.size == aug_data.size, "data size is not the same"
def test_time_masking(logger: Logger = None):
"""
Checks that the number of time masks are less than the maximum number allowed.
Values of test_tuples are:
('time_warping_para', 'frequency_masking_para', 'time_masking_para'
'frequency_mask_num', 'time_mask_num')
"""
test_tuples = [
(0, 0, 60, 0, 1), # 1 mask with max width of 60
(0, 0, 30, 0, 2),
(0, 0, 20, 0, 3)
]
audio_paths = get_all_test_audio()
number_of_tests = 10 # multiple tests as mask selection is random
for _ in range(number_of_tests):
for audio_path in audio_paths:
for param_tuple in test_tuples:
audio_data, samp_rate = array_from_wave(audio_path)
features = log_spectrogram_from_data(audio_data,
samp_rate,
window_size=32,
step_size=16)
features = torch.from_numpy(features.T)
aug_features = spec_augment(features, *param_tuple)
aug_features = to_numpy(aug_features)
num_mask_rows = count_time_mask(aug_features)
time_mask_size = param_tuple[2]
num_time_masks = param_tuple[4]
max_time_masks = time_mask_size * num_time_masks
#print(f"number of time masked rows: {num_mask_rows}, max_time_masked: {max_time_masks}")
assert num_mask_rows <= max_time_masks
def test_datatype():
test_audio_paths = get_all_test_audio()
snr_level = 30
for audio_path in test_audio_paths:
audio_data, sr = array_from_wave(audio_path)
augmented_data = synthetic_gaussian_noise_inject(audio_data,
snr_range=(snr_level,
snr_level))
assert augmented_data.dtype == "int16"
def test_audio_feature_normalize():
"""
"""
audio_files = get_all_test_audio()
for audio_file in audio_files:
feature = log_spectrogram_from_file(audio_file)
normalized_feature = feature_normalize(feature)
np.testing.assert_almost_equal(normalized_feature.mean(), 0)
np.testing.assert_almost_equal(normalized_feature.std(), 1)
def test_apply_spec_augment_call(logger: Logger = None):
"""
Just tests if the apply_spec_augment can be called without errors
Arguments:
logger - Logger: can be taken as input to teset logger
"""
audio_paths = get_all_test_audio()
for audio_path in audio_paths:
audio_data, samp_rate = array_from_wave(audio_path)
features = log_spectrogram_from_data(audio_data,
samp_rate,
window_size=32,
step_size=16)
apply_spec_augment(features, logger)
def test_high_snr_value():
test_audio_paths = get_all_test_audio()
snr_level = 100
# absolute tolerance is 1e-5 of the range of values in pcm16 format (2**16)
atol = 2**16 * 1e-5
for audio_path in test_audio_paths:
audio_data, sr = array_from_wave(audio_path)
augmented_data = synthetic_gaussian_noise_inject(audio_data,
snr_range=(snr_level,
snr_level))
np.testing.assert_allclose(audio_data,
augmented_data,
rtol=1e-03,
atol=atol)
def test_no_augment():
"""
tests that input audio and augmented data are identical with no augmentation: tempo=1.0, gain=0
pitch = 0
"""
tempo = 1.0
gain = 0.0
pitch = 0.0
audio_paths = get_all_test_audio()
for audio_path in audio_paths:
# un-augmented audio_data
audio_data, samp_rate = array_from_wave(audio_path)
aug_data, samp_rate = tempo_gain_pitch_perturb(audio_path,
sample_rate=samp_rate,
tempo_range=(tempo,
tempo),
gain_range=(gain, gain),
pitch_range=(pitch,
pitch))
assert all(audio_data == aug_data), "data is not the same"
def test_gain_increase_amplitude():
"""
tests that 1) 6 dB increase in gain coorespondes to a 1.995 increase in the sum of the absolute
value of the amplitudes and,
2) a 6 db decrease cooresponds to a 0.5 decrease in the sum abs value of amplitudes
Ratio is computed as: ratio = 10**(gain/20)
"""
tempo = 1.0
pitch = 0.0
gain_ratio_tuples = [
(0, 1.0), # not augmentation
(6, 1.995),
(-6, 0.501)
]
#(10, 3.162), # these two tests fail.
#(-10, 0.3162) # I'm not sure why, likely an error in my approach.
audio_paths = get_all_test_audio() # only using a single audio path
for audio_path in audio_paths:
print(f"audio_path: {audio_path}")
for gain, amp_ratio in gain_ratio_tuples:
# un-augmented audio_data
audio_data, samp_rate = array_from_wave(audio_path)
aug_data, samp_rate = tempo_gain_pitch_perturb(
audio_path,
sample_rate=samp_rate,
tempo_range=(tempo, tempo),
gain_range=(gain, gain),
pitch_range=(pitch, pitch))
audio_rms = audioop.rms(audio_data, 2)
scaled_aug_rms = audioop.rms(aug_data, 2) / amp_ratio
accuracy = -1 # same up to 10^(-accuracy)
print(
f"audio rms: {audio_rms}, scaled_aug rms: {scaled_aug_rms}, ratio:{amp_ratio}, accuracy:{10**(-accuracy)}"
)
np.testing.assert_almost_equal(audio_rms,
scaled_aug_rms,
decimal=accuracy)
def test_tempo_augment():
"""
Verifies the size of the augmented data scaled by the tempo equals the size
of the un-augmented data
"""
audio_paths = get_all_test_audio()
tempos = [0, 0.5, 0.85, 1, 1.15, 2]
for audio_path in audio_paths:
# un-augmented audio_data
audio_data, samp_rate = array_from_wave(audio_path)
for tempo in tempos:
aug_data, samp_rate = tempo_gain_pitch_perturb(
audio_path,
sample_rate=samp_rate,
tempo_range=(tempo, tempo),
gain_range=(0, 0),
pitch_range=(0, 0))
print(
f"audio_data size: {audio_data.size}, aug_data: {aug_data.size}, tempo: {tempo}"
)
assert audio_data.size == pytest.approx(aug_data.size * tempo, 1e-1)