def test_vad(self):
common_utils.set_audio_backend('default')
filepath = common_utils.get_asset_path("vad-go-mono-32000.wav")
waveform, sample_rate = torchaudio.load(filepath)
self.assert_batch_consistencies(F.vad,
waveform,
sample_rate=sample_rate)
def assert_compatibilities(self, n_fft, hop_length, power, n_mels, n_mfcc,
sample_rate):
common_utils.set_audio_backend('default')
path = common_utils.get_asset_path('sinewave.wav')
sound, sample_rate = common_utils.load_wav(path)
sound_librosa = sound.cpu().numpy().squeeze() # (64000)
# test core spectrogram
spect_transform = torchaudio.transforms.Spectrogram(
n_fft=n_fft, hop_length=hop_length, power=power)
out_librosa, _ = librosa.core.spectrum._spectrogram(
y=sound_librosa, n_fft=n_fft, hop_length=hop_length, power=power)
out_torch = spect_transform(sound).squeeze().cpu()
self.assertEqual(out_torch,
torch.from_numpy(out_librosa),
atol=1e-5,
rtol=1e-5)
# test mel spectrogram
melspect_transform = torchaudio.transforms.MelSpectrogram(
sample_rate=sample_rate,
window_fn=torch.hann_window,
hop_length=hop_length,
n_mels=n_mels,
n_fft=n_fft)
librosa_mel = librosa.feature.melspectrogram(y=sound_librosa,
sr=sample_rate,
n_fft=n_fft,
hop_length=hop_length,
n_mels=n_mels,
htk=True,
norm=None)
librosa_mel_tensor = torch.from_numpy(librosa_mel)
torch_mel = melspect_transform(sound).squeeze().cpu()
self.assertEqual(torch_mel.type(librosa_mel_tensor.dtype),
librosa_mel_tensor,
atol=5e-3,
rtol=1e-5)
# test s2db
power_to_db_transform = torchaudio.transforms.AmplitudeToDB(
'power', 80.)
power_to_db_torch = power_to_db_transform(
spect_transform(sound)).squeeze().cpu()
power_to_db_librosa = librosa.core.spectrum.power_to_db(out_librosa)
self.assertEqual(power_to_db_torch,
torch.from_numpy(power_to_db_librosa),
atol=5e-3,
rtol=1e-5)
mag_to_db_transform = torchaudio.transforms.AmplitudeToDB(
'magnitude', 80.)
mag_to_db_torch = mag_to_db_transform(torch.abs(sound)).squeeze().cpu()
mag_to_db_librosa = librosa.core.spectrum.amplitude_to_db(
sound_librosa)
self.assertEqual(mag_to_db_torch,
torch.from_numpy(mag_to_db_librosa),
atol=5e-3,
rtol=1e-5)
power_to_db_torch = power_to_db_transform(
melspect_transform(sound)).squeeze().cpu()
db_librosa = librosa.core.spectrum.power_to_db(librosa_mel)
db_librosa_tensor = torch.from_numpy(db_librosa)
self.assertEqual(power_to_db_torch.type(db_librosa_tensor.dtype),
db_librosa_tensor,
atol=5e-3,
rtol=1e-5)
# test MFCC
melkwargs = {'hop_length': hop_length, 'n_fft': n_fft}
mfcc_transform = torchaudio.transforms.MFCC(sample_rate=sample_rate,
n_mfcc=n_mfcc,
norm='ortho',
melkwargs=melkwargs)
# librosa.feature.mfcc doesn't pass kwargs properly since some of the
# kwargs for melspectrogram and mfcc are the same. We just follow the
# function body in
# https://librosa.github.io/librosa/_modules/librosa/feature/spectral.html#melspectrogram
# to mirror this function call with correct args:
#
# librosa_mfcc = librosa.feature.mfcc(
# y=sound_librosa, sr=sample_rate, n_mfcc = n_mfcc,
# hop_length=hop_length, n_fft=n_fft, htk=True, norm=None, n_mels=n_mels)
librosa_mfcc = scipy.fftpack.dct(db_librosa,
axis=0,
type=2,
norm='ortho')[:n_mfcc]
librosa_mfcc_tensor = torch.from_numpy(librosa_mfcc)
torch_mfcc = mfcc_transform(sound).squeeze().cpu()
self.assertEqual(torch_mfcc.type(librosa_mfcc_tensor.dtype),
librosa_mfcc_tensor,
atol=5e-3,
rtol=1e-5)