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 signal_augmentations(self, wave_file: str) -> tuple:
"""
Performs all of the augmtations to the raw audio signal. The audio data is in pcm16 format.
Arguments:
wave_file - str: the path to the audio sample
Returns:
audio_data - np.ndarray: augmented np-array
samp_rate - int: sample rate of the audio recording
"""
if self.use_log:
self.logger.info(f"preproc: audio_data read: {wave_file}")
audio_data, samp_rate = array_from_wave(wave_file)
# sox-based tempo, gain, pitch augmentations
if self.tempo_gain_pitch_perturb and self.train_status:
if np.random.binomial(1, self.tempo_gain_pitch_prob):
audio_data, samp_rate = tempo_gain_pitch_perturb(
wave_file,
samp_rate,
self.tempo_range,
self.gain_range,
self.pitch_range,
self.augment_from_normal,
logger=self.logger)
if self.use_log:
self.logger.info(f"preproc: tempo_gain_pitch applied")
# synthetic gaussian noise
if self.synthetic_gaussian_noise and self.train_status:
if np.random.binomial(1, self.gauss_noise_prob):
audio_data = synthetic_gaussian_noise_inject(
audio_data,
self.gauss_snr_db_range,
self.augment_from_normal,
logger=self.logger)
if self.use_log:
self.logger.info(f"preproc: synth_gauss_noise applied")
# noise injection
if self.background_noise and self.train_status:
if np.random.binomial(1, self.background_noise_prob):
audio_data = inject_noise(audio_data, samp_rate,
self.noise_dir,
self.background_noise_range,
self.augment_from_normal,
self.logger)
if self.use_log: self.logger.info(f"preproc: noise injected")
return audio_data, samp_rate
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_regression_equal_pickle():
"""
The pickle data is output from using the Speak-out.wav file with an snr_level = 30 and a random seed of zero
"""
pickle_path = "../test_pickle/sythentic-gaussian-noise-inject_Speak-out_snr-30.pickle"
with open(pickle_path, 'rb') as fid:
pickle_data = pickle.load(fid)
audio_path = "../test_audio/Speak-out.wav"
snr_level = 30
audio_data, sr = array_from_wave(audio_path)
np.random.seed(0)
augmented_data = synthetic_gaussian_noise_inject(audio_data,
snr_range=(snr_level,
snr_level))
assert (augmented_data == pickle_data
).sum() == augmented_data.size, "regression test fails"
def test_float_input_failure():
float_data = np.arange(100, dtype=np.float64)
snr_level = 30
with pytest.raises(AssertionError) as execinfo:
synthetic_gaussian_noise_inject(float_data,
snr_range=(snr_level, snr_level))
def test_zero_input():
empty_data = np.empty((0, ), dtype="int16")
snr_level = 30
synthetic_gaussian_noise_inject(empty_data,
snr_range=(snr_level, snr_level))