def test_n_output_audio(self):
audio, sampling_rate = AudioLoader.load_audio(self.sample_wav_file)
flows = [
naf.Sequential(
[naa.CropAug(sampling_rate=sampling_rate),
naa.LoudnessAug()]),
naf.Sometimes(
[naa.CropAug(sampling_rate=sampling_rate),
naa.LoudnessAug()],
pipeline_p=0.9),
naf.Sequential([
naf.Sequential([
naa.CropAug(sampling_rate=sampling_rate),
naa.LoudnessAug()
]),
naf.Sometimes([
naa.CropAug(sampling_rate=sampling_rate),
naa.LoudnessAug()
],
pipeline_p=0.9)
])
]
for flow in flows:
augmented_audios = flow.augment(audio, n=3)
self.assertGreater(len(augmented_audios), 1)
for augmented_audio in augmented_audios:
self.assertFalse(np.array_equal(audio, augmented_audio))
self.assertLess(0, len(flows))
def test_substitute(self):
aug = naa.LoudnessAug()
augmented_audio = aug.augment(self.audio)
self.assertFalse(np.array_equal(self.audio, augmented_audio))
self.assertEqual(len(self.audio), len(augmented_audio))
self.assertTrue(self.sampling_rate > 0)
def augment_audio(data, sampling_rate, loudness_factor, speed_range, pitch_range, shift_max, noise_factor):
"""
OPTIONS:
LOUDNESS AUGMENTER: naa.LoudnessAug()
MASK AUGMENTER: naa.MaskAug()
SPEED AUGMENTER: naa.SpeedAug()
SHIFT AUGMENTER: naa.ShiftAug()
NOISE AUGMENTER: naa.NoiseAug()
CROP AUGMENTER: naa.CropAug()
PITCH AUGMENTER_ naa.PitchAug()
"""
flow = naf.Sequential([naa.LoudnessAug(loudness_factor),
naa.SpeedAug(speed_range),
naa.PitchAug(sampling_rate = sampling_rate, pitch_range = pitch_range),
naa.ShiftAug(sampling_rate = sampling_rate, shift_max = shift_max),
naa.NoiseAug(noise_factor),
])
augmented_audio = flow.augment(data)
return augmented_audio
def test_substitute(self):
audio, sampling_rate = librosa.load(self.sample_wav_file)
aug = naa.LoudnessAug()
augmented_audio = aug.augment(audio)
self.assertFalse(np.array_equal(audio, augmented_audio))
self.assertEqual(len(audio), len(augmented_audio))
self.assertTrue(sampling_rate > 0)
def gen_augmented(original, sample_rate):
# list of augmentors available from the nlpaug library
augment_list = [
#naa.CropAug(sampling_rate=sample_rate)
naa.NoiseAug(),
naa.SpeedAug(),
naa.LoudnessAug(factor=(0.5, 2)),
naa.VtlpAug(sampling_rate=sample_rate, zone=(0.0, 1.0)),
naa.PitchAug(sampling_rate=sample_rate, factor=(-1, 3))
]
# sample augmentation randomly
aug_idx = random.randint(0, len(augment_list) - 1)
augmented_data = augment_list[aug_idx].augment(original)
return augmented_data
def process_audio_files_with_aug(filename, dirpath):
sr = 16000
audio_array, sample_rate = librosa.load(os.path.join(dirpath, 'flac', filename), sr=sr)
aug_crop = naa.CropAug(sampling_rate=sr)
audio_array_crop = aug_crop.augment(audio_array)
aug_loud = naa.LoudnessAug(loudness_factor=(2, 5))
audio_array_loud = aug_loud.augment(audio_array)
aug_noise = naa.NoiseAug(noise_factor=0.03)
audio_array_noise = aug_noise.augment(audio_array)
audio_array_list= [audio_array,audio_array_crop,audio_array_loud,
audio_array_noise ]
out_list = convert_audio_to_processed_list(audio_array_list, filename, dirpath)
return out_list
def test_zone(self):
zone = (0, 1)
coverage = 1
aug = naa.LoudnessAug(zone=zone, coverage=coverage)
aug.model.stateless = False
augmented_audio = aug.augment(self.audio)
reconstruct_augmented_audio = np.concatenate(
(self.audio[:aug.model.start_pos], aug.model.aug_data,
self.audio[aug.model.end_pos:]),
axis=0)
self.assertTrue(
np.array_equal(augmented_audio, reconstruct_augmented_audio))
self.assertTrue(len(aug.model.aug_data),
int(len(self.audio) * (zone[1] - zone[0]) * coverage))
def process_audio_files_with_aug(self, audio_path):
sr = 16000
audio_array, sample_rate = librosa.load(audio_path, sr=sr)
aug_crop = naa.CropAug(sampling_rate=sr)
audio_array_crop = aug_crop.augment(audio_array)
aug_loud = naa.LoudnessAug(loudness_factor=(2, 5))
audio_array_loud = aug_loud.augment(audio_array)
aug_noise = naa.NoiseAug(noise_factor=0.03)
audio_array_noise = aug_noise.augment(audio_array)
mel_spec_array_load = melspectrogram(audio_array_loud, hparams=hparams)
mel_spec_array_noise = melspectrogram(audio_array_noise,
hparams=hparams)
audio_array_list = [mel_spec_array_load, mel_spec_array_noise]
# audio_array_list= [audio_array_crop,audio_array_loud,
# audio_array_noise ]
return audio_array_list
def test_coverage_and_zone(self):
params = [
((0.3, 0.7), 1),
((0, 1), 1)
]
for zone, coverage in params:
augs = [
naa.LoudnessAug(zone=zone, coverage=coverage, stateless=False),
naa.MaskAug(zone=zone, coverage=coverage, stateless=False),
naa.NoiseAug(zone=zone, coverage=coverage, stateless=False),
naa.PitchAug(zone=zone, coverage=coverage, stateless=False, sampling_rate=self.sampling_rate),
naa.SpeedAug(zone=zone, coverage=coverage, stateless=False),
naa.VtlpAug(zone=zone, coverage=coverage, stateless=False, sampling_rate=self.sampling_rate),
naa.NormalizeAug(zone=zone, coverage=coverage, stateless=False),
naa.PolarityInverseAug(zone=zone, coverage=coverage, stateless=False)
]
for aug in augs:
aug_data = aug.augment(self.audio)
self.assertTrue(len(aug_data[aug.start_pos:aug.end_pos]), int(len(self.audio) * (zone[1] - zone[0]) * coverage))
def test_empty_input(self):
audio = np.array([])
aug = naa.LoudnessAug()
augmented_audio = aug.augment(audio)
self.assertTrue(np.array_equal(audio, augmented_audio))
def aug_loudness(data, loudness_factor=(0.95, 1.5)): aug = naa.LoudnessAug(loudness_factor=loudness_factor) augmented_data = aug.augment(data) return augmented_data
def loudness(data):
aug = naa.LoudnessAug()
return aug.augment(data)
def get_random_loudness(self):
return naa.LoudnessAug(crop=(0, 1), coverage=1)
