def __getitem__(self, idx):
if self.feature_type == "waveform":
path, label = self.item[idx]
return (load_wave(path), label)
else:
path, label, trans, utt = self.item[idx]
feat = kaldiio.load_mat(path)
tmp = "aug"
if tmp == "aug" and self.train:
feat = spec_augment(feat)
# TODO
trans = [data_enhancement(tran) for tran in trans]
trans = sum(trans, [])
feat = skip_feat(make_context(feat, self.left_ctx, self.right_ctx),
self.n_skip_frame)
seq_len, dim = feat.shape
if seq_len % self.n_downsample != 0:
pad_len = self.n_downsample - seq_len % self.n_downsample
feat = np.vstack([feat, np.zeros((pad_len, dim))])
if self.mel:
return (F_Mel(torch.from_numpy(feat), audio_conf), label)
else:
return (torch.from_numpy(feat), torch.LongTensor(label),
torch.LongTensor(trans), utt)
def __getitem__(self, idx):
path, utt = self.item[idx]
feat = kaldiio.load_mat(path)
feat = skip_feat(make_context(feat, self.left_ctx, self.right_ctx),
self.n_skip_frame)
seq_len, dim = feat.shape
if seq_len % self.n_downsample != 0:
pad_len = self.n_downsample - seq_len % self.n_downsample
feat = np.vstack([feat, np.zeros((pad_len, dim))])
if self.mel:
return F_Mel(torch.from_numpy(feat), audio_conf)
else:
return (torch.from_numpy(feat), utt)
def __getitem__(self, idx):
if self.feature_type == "waveform":
path, label = self.item[idx]
return (load_wave(path), label)
else:
path, label, utt = self.item[idx]
feat = kaldiio.load_mat(path)
feat = skip_feat(make_context(feat, self.left_ctx, self.right_ctx),
self.n_skip_frame)
seq_len, dim = feat.shape
if seq_len % self.n_downsample != 0:
pad_len = self.n_downsample - seq_len % self.n_downsample
feat = np.vstack([feat, np.zeros((pad_len, dim))])
if self.mel:
return (F_Mel(torch.from_numpy(feat), audio_conf), label)
else:
return (torch.from_numpy(feat), torch.LongTensor(label), utt)
def __getitem__(self, idx):
if self.feature_type == "waveform":
path, label = self.item[idx]
return (load_wave(path), label)
else:
path, label, trans, utt = self.item[idx]
feat = kaldiio.load_mat(path)
## data agument
if (self.is_training == True):
feat = spec_augment(feat)
trans = [data_enhance(phoneme) for phoneme in trans]
feat = skip_feat(make_context(feat, self.left_ctx, self.right_ctx),
self.n_skip_frame)
seq_len, dim = feat.shape
if seq_len % self.n_downsample != 0:
pad_len = self.n_downsample - seq_len % self.n_downsample
feat = np.vstack([feat, np.zeros((pad_len, dim))])
if self.mel:
return (F_Mel(torch.from_numpy(feat), audio_conf), label)
else:
return (torch.from_numpy(feat), torch.LongTensor(label),
torch.LongTensor(trans), utt)