def test_fill_zeros(self):
samples, sample_rate = load_sound_file(
os.path.join(DEMO_DIR, "acoustic_guitar_0.wav"), sample_rate=None
)
magnitude_spectrogram = librosa.feature.melspectrogram(
y=samples, sr=sample_rate
)
mask_fraction = 0.05
transform = SpecFrequencyMask(
fill_mode="constant",
fill_constant=0.0,
min_mask_fraction=mask_fraction,
max_mask_fraction=mask_fraction,
p=1.0,
)
augmented_spectrogram = transform(magnitude_spectrogram)
if DEBUG:
plot_matrix(np.log(augmented_spectrogram))
with np.testing.assert_raises(AssertionError):
np.testing.assert_array_equal(augmented_spectrogram, magnitude_spectrogram)
num_zeroed_frequencies = 0
for i in range(augmented_spectrogram.shape[0]):
if sum(augmented_spectrogram[i]) == 0.0:
num_zeroed_frequencies += 1
self.assertEqual(
num_zeroed_frequencies,
int(round(magnitude_spectrogram.shape[0] * mask_fraction)),
)
def test_fill_mean_multichannel(self):
samples, sample_rate = load_sound_file(
os.path.join(DEMO_DIR, "background_noises", "hens.ogg"),
sample_rate=None,
mono=False,
)
assert samples.shape[0] == 2
magnitude_spectrogram_chn0 = librosa.feature.melspectrogram(
y=np.asfortranarray(samples[0, :]), sr=sample_rate
)
magnitude_spectrogram_chn1 = librosa.feature.melspectrogram(
y=np.asfortranarray(samples[1, :]), sr=sample_rate
)
multichannel_magnitude_spectrogram = np.zeros(
shape=(
magnitude_spectrogram_chn0.shape[0],
magnitude_spectrogram_chn0.shape[1],
3,
),
dtype=np.float32,
)
multichannel_magnitude_spectrogram[:, :, 0] = magnitude_spectrogram_chn0
multichannel_magnitude_spectrogram[:, :, 1] = magnitude_spectrogram_chn1
multichannel_magnitude_spectrogram[:, :, 2] = magnitude_spectrogram_chn1
mask_fraction = 0.05
transform = SpecFrequencyMask(
fill_mode="mean",
min_mask_fraction=mask_fraction,
max_mask_fraction=mask_fraction,
p=1.0,
)
augmented_spectrogram = transform(multichannel_magnitude_spectrogram)
if DEBUG:
image = (7 + np.log10(augmented_spectrogram + 0.0000001)) / 8
plot_matrix(image)
with np.testing.assert_raises(AssertionError):
np.testing.assert_array_equal(
augmented_spectrogram, multichannel_magnitude_spectrogram
)
num_masked_frequencies = 0
for i in range(augmented_spectrogram.shape[0]):
frequency_slice = augmented_spectrogram[i]
if (
np.amin(frequency_slice) == np.amax(frequency_slice)
and np.sum(frequency_slice) != 0.0
):
num_masked_frequencies += 1
self.assertEqual(
num_masked_frequencies,
int(round(multichannel_magnitude_spectrogram.shape[0] * mask_fraction)),
)
def test_shuffle_channels_mono(self):
samples, sample_rate = load_sound_file(os.path.join(
DEMO_DIR, "acoustic_guitar_0.wav"),
sample_rate=None)
magnitude_spectrogram = librosa.feature.melspectrogram(y=samples,
sr=sample_rate)
transform = SpecChannelShuffle(p=1.0)
with self.assertRaises(MonoAudioNotSupportedException):
augmented_spectrogram = transform(magnitude_spectrogram)
def test_shuffle_channels(self):
samples, sample_rate = load_sound_file(
os.path.join(DEMO_DIR, "background_noises", "hens.ogg"),
sample_rate=None,
mono=False,
)
assert samples.shape[0] == 2
magnitude_spectrogram_chn0 = librosa.feature.melspectrogram(
y=np.asfortranarray(samples[0, :]), sr=sample_rate)
magnitude_spectrogram_chn1 = librosa.feature.melspectrogram(
y=np.asfortranarray(samples[1, :]), sr=sample_rate)
multichannel_magnitude_spectrogram = np.zeros(
shape=(
magnitude_spectrogram_chn0.shape[0],
magnitude_spectrogram_chn0.shape[1],
3,
),
dtype=np.float32,
)
multichannel_magnitude_spectrogram[:, :,
0] = magnitude_spectrogram_chn0
multichannel_magnitude_spectrogram[:, :,
1] = magnitude_spectrogram_chn1
multichannel_magnitude_spectrogram[:, :,
2] = magnitude_spectrogram_chn1 * 0.7
if DEBUG:
image = (7 + np.log10(multichannel_magnitude_spectrogram +
0.0000001)) / 8
plot_matrix(image, title="before")
# Make the shuffled channels do not equal the original order
transform = SpecChannelShuffle(p=1.0)
for _ in range(100000):
transform.randomize_parameters(multichannel_magnitude_spectrogram)
if transform.parameters["shuffled_channel_indexes"] != [0, 1, 2]:
break
transform.freeze_parameters()
augmented_spectrogram = transform(multichannel_magnitude_spectrogram)
if DEBUG:
image = (7 + np.log10(augmented_spectrogram + 0.0000001)) / 8
plot_matrix(image, title="after")
with np.testing.assert_raises(AssertionError):
np.testing.assert_array_equal(augmented_spectrogram,
multichannel_magnitude_spectrogram)
for augmented_index, original_index in enumerate(
transform.parameters.get("shuffled_channel_indexes")):
np.testing.assert_array_equal(
augmented_spectrogram[:, :, augmented_index],
multichannel_magnitude_spectrogram[:, :, original_index],
)
def test_load_stereo_signed_16_bit_wav(self):
samples, sample_rate = load_sound_file(os.path.join(
DEMO_DIR, "stereo_16bit.wav"),
sample_rate=None)
self.assertEqual(sample_rate, 16000)
self.assertEqual(samples.dtype, np.float32)
self.assertEqual(len(samples.shape), 1)
self.assertEqual(samples.shape[0], 17833)
max_value = np.amax(samples)
self.assertGreater(max_value, 0.5)
self.assertLess(max_value, 1.0)
def test_load_mono_ms_adpcm(self):
samples, sample_rate = load_sound_file(os.path.join(
DEMO_DIR, "ms_adpcm.wav"),
sample_rate=None)
self.assertEqual(sample_rate, 11024)
self.assertEqual(samples.dtype, np.float32)
self.assertEqual(len(samples.shape), 1)
self.assertEqual(samples.shape[0], 895500)
max_value = np.amax(samples)
self.assertGreater(max_value, 0.3)
self.assertLess(max_value, 1.0)
def test_load_mono_signed_24_bit_wav(self):
samples, sample_rate = load_sound_file(os.path.join(
DEMO_DIR, "signed_24bit.wav"),
sample_rate=None)
self.assertEqual(sample_rate, 48000)
self.assertEqual(samples.dtype, np.float32)
self.assertEqual(len(samples.shape), 1)
self.assertEqual(samples.shape[0], 54514)
max_value = np.amax(samples)
self.assertGreater(max_value, 0.09)
self.assertLess(max_value, 1.0)
def test_load_mono_m4a(self):
samples, sample_rate = load_sound_file(os.path.join(
DEMO_DIR, "testing.m4a"),
sample_rate=None)
self.assertEqual(sample_rate, 44100)
self.assertEqual(samples.dtype, np.float32)
self.assertEqual(len(samples.shape), 1)
self.assertGreaterEqual(samples.shape[0], 141312)
self.assertLessEqual(samples.shape[0], 141312)
max_value = np.amax(samples)
self.assertGreater(max_value, 0.1)
self.assertLess(max_value, 1.0)
def test_load_mono_opus(self):
samples, sample_rate = load_sound_file(os.path.join(
DEMO_DIR, "bus.opus"),
sample_rate=None)
self.assertEqual(samples.dtype, np.float32)
self.assertEqual(len(samples.shape), 1)
# Apparently, the exact duration may vary slightly based on which decoder is used
self.assertGreaterEqual(samples.shape[0], 36682)
self.assertLessEqual(samples.shape[0], 36994)
max_value = np.amax(samples)
self.assertGreater(max_value, 0.3)
self.assertLess(max_value, 1.0)
def test_load_stereo_ogg_vorbis(self):
samples, sample_rate = load_sound_file(os.path.join(
DEMO_DIR, "background_noises", "hens.ogg"),
sample_rate=None)
self.assertEqual(samples.dtype, np.float32)
self.assertEqual(len(samples.shape), 1)
# Apparently, the exact duration may vary slightly based on which decoder is used
self.assertGreaterEqual(samples.shape[0], 442575)
self.assertLessEqual(samples.shape[0], 443328)
max_value = np.amax(samples)
self.assertGreater(max_value, 0.02)
self.assertLess(max_value, 1.0)
def test_fill_mean(self):
samples, sample_rate = load_sound_file(
os.path.join(DEMO_DIR, "acoustic_guitar_0.wav"), sample_rate=None
)
magnitude_spectrogram = librosa.feature.melspectrogram(
y=samples, sr=sample_rate
)
min_mask_fraction = 0.05
max_mask_fraction = 0.09
transform = SpecFrequencyMask(
fill_mode="mean",
min_mask_fraction=min_mask_fraction,
max_mask_fraction=max_mask_fraction,
p=1.0,
)
augmented_spectrogram = transform(magnitude_spectrogram)
if DEBUG:
plot_matrix(np.log(augmented_spectrogram))
num_masked_frequencies = 0
for i in range(augmented_spectrogram.shape[0]):
frequency_slice = augmented_spectrogram[i]
if (
np.amin(frequency_slice) == np.amax(frequency_slice)
and sum(frequency_slice) != 0.0
):
num_masked_frequencies += 1
self.assertGreaterEqual(
num_masked_frequencies,
int(round(magnitude_spectrogram.shape[0] * min_mask_fraction)),
)
self.assertLessEqual(
num_masked_frequencies,
int(round(magnitude_spectrogram.shape[0] * max_mask_fraction)),
)
def test_load_mono_ms_adpcm_and_resample(self):
with warnings.catch_warnings(record=True) as w:
# Cause all warnings to always be triggered.
warnings.simplefilter("always")
samples, sample_rate = load_sound_file(os.path.join(
DEMO_DIR, "ms_adpcm.wav"),
sample_rate=16000)
assert len(w) == 1
assert (
"resampled from 11024 hz to 16000 hz. This hurt execution time"
in str(w[-1].message))
self.assertEqual(sample_rate, 16000)
self.assertEqual(samples.dtype, np.float32)
self.assertEqual(len(samples.shape), 1)
self.assertEqual(samples.shape[0], math.ceil(895500 * 16000 / 11024))
max_value = np.amax(samples)
self.assertGreater(max_value, 0.3)
self.assertLess(max_value, 1.0)
"instance": TimeMask(p=1.0),
"num_runs": 5
},
{
"instance": TimeStretch(min_rate=0.8, max_rate=1.25, p=1.0),
"num_runs": 5
},
{
"instance": Trim(p=1.0),
"num_runs": 1
},
]
for sound_file_path in sound_file_paths:
samples, sample_rate = load_sound_file(sound_file_path,
sample_rate=None,
mono=False)
if len(samples.shape) == 2 and samples.shape[0] > samples.shape[1]:
samples = samples.transpose()
print("Transforming {} with shape {}".format(sound_file_path.name,
str(samples.shape)))
execution_times = {}
for transform in transforms:
augmenter = transform["instance"]
run_name = (transform.get("name") if transform.get("name") else
transform["instance"].__class__.__name__)
execution_times[run_name] = []
for i in range(transform["num_runs"]):
output_file_path = os.path.join(
def __load_sound(file_path, sample_rate):
return load_sound_file(file_path, sample_rate)
