def __getitem__(self, index):
wavpath = self.wavpaths[index]
wav, sr = torchaudio.load(wavpath)
wav = wav.squeeze().numpy()
if self.split == 'train':
transforms = WavformAugmentation(sr)
wav = transforms(wav)
# pad to 150k frames
if len(wav) > self.max_length:
# randomly pick which side to chop off (fix if validation)
flip = (bool(random.getrandbits(1))
if self.split == 'train' else True)
padded = (wav[:self.max_length]
if flip else wav[-self.max_length:])
else:
padded = np.zeros(self.max_length)
padded[:len(wav)] = wav # pad w/ silence
label = self.labels[index]
if self.caller_intent == 'dialog_acts':
label = torch.LongTensor(label)
elif self.caller_intent == 'sentiment':
label = [label['positive'], label['neutral'], label['negative']]
label = torch.FloatTensor(label)
padded = torch.from_numpy(padded).float()
return index, padded, label
def __getitem__(self, index):
wav_name = self.wav_paths[index]
wav_path = os.path.join(self.root, wav_name)
label = self.labels[index]
if self.label_type == 'action':
label = FLUENTSPEECH_ACTIONS.index(label)
elif self.label_type == 'object':
label = FLUENTSPEECH_OBJECTS.index(label)
elif self.label_type == 'location':
label = FLUENTSPEECH_LOCATIONS.index(label)
wavform, sample_rate = torchaudio.load(wav_path)
wavform = wavform[0].numpy()
if self.wavform_transforms:
transforms = WavformAugmentation(sample_rate)
wavform = transforms(wavform)
# pad to 150k frames
if len(wavform) > self.max_length:
# randomly pick which side to chop off (fix if validation)
flip = (bool(random.getrandbits(1)) if self.train else True)
padded = (wavform[:self.max_length]
if flip else wavform[-self.max_length:])
else:
padded = np.zeros(self.max_length)
padded[:len(wavform)] = wavform # pad w/ silence
hop_length_dict = {224: 672, 112: 1344, 64: 2360, 32: 4800}
spectrum = librosa.feature.melspectrogram(
padded,
sample_rate,
hop_length=hop_length_dict[self.input_size],
n_mels=self.input_size,
)
if self.spectral_transforms: # apply time and frequency masks
transforms = SpectrumAugmentation()
spectrum = transforms(spectrum)
# log mel-spectrogram
spectrum = librosa.power_to_db(spectrum**2)
spectrum = torch.from_numpy(spectrum).float()
spectrum = spectrum.unsqueeze(0)
if self.spectral_transforms: # apply noise on spectral
noise_stdev = 0.25 * self.normalize_stdev[0]
noise = torch.randn_like(spectrum) * noise_stdev
spectrum = spectrum + noise
normalize = Normalize(self.normalize_mean, self.normalize_stdev)
spectrum = normalize(spectrum)
return index, spectrum, int(label)
def __getitem__(self, index):
wavpath = self.wavpaths[index]
wav, sr = torchaudio.load(wavpath)
wav = wav.squeeze().numpy()
if self.wavform_transforms and self.split == 'train':
transforms = WavformAugmentation(sr)
wav = transforms(wav)
# pad to 150k frames
if len(wav) > self.max_length:
# randomly pick which side to chop off (fix if validation)
flip = (bool(random.getrandbits(1))
if self.split == 'train' else True)
padded = (wav[:self.max_length]
if flip else wav[-self.max_length:])
else:
padded = np.zeros(self.max_length)
padded[:len(wav)] = wav # pad w/ silence
spectrum = librosa.feature.melspectrogram(
padded,
sr,
hop_length=HARPER_VALLEY_HOP_LENGTH_DICT[self.input_size],
n_mels=self.input_size,
)
if self.spectral_transforms: # apply time and frequency masks
transforms = SpectrumAugmentation()
spectrum = transforms(spectrum)
# log mel-spectrogram
spectrum = librosa.power_to_db(spectrum**2)
spectrum = torch.from_numpy(spectrum).float()
spectrum = spectrum.unsqueeze(0)
if self.spectral_transforms: # apply noise on spectral
noise_stdev = 0.25 * self.normalize_stdev[0]
noise = torch.randn_like(spectrum) * noise_stdev
spectrum = spectrum + noise
label = self.labels[index]
if self.caller_intent == 'dialog_acts':
label = torch.LongTensor(label)
elif self.caller_intent == 'sentiment':
label = [label['positive'], label['neutral'], label['negative']]
label = torch.FloatTensor(label)
normalize = Normalize(self.normalize_mean, self.normalize_stdev)
spectrum = normalize(spectrum)
return index, spectrum, label