def __getitem__(self, idx):
uttname = self.uttlist[idx]
info_filename = "{}/data/{}.txt".format(self.data_dir, uttname)
assert os.path.exists(info_filename)
with open(info_filename, 'r') as fh:
info = fh.readline().strip('\n')
info_split = info.split(None, 3)
# each record has 4 fields
# (1)uttname (2)uttdur (3)label file (4)feature file
feat_file, label_file = info_split[3], info_split[2]
# compute STFT feature
data, samplerate = kaldi_data.load_wav(feat_file)
Y = feature.stft(data, self.frame_size, self.frame_shift)
feat = feature.transform(Y, self.input_transform)
# prepare diarization label
label = self.process_label_file(label_file)
second_per_frame = self.frame_shift * 1.0 / self.rate
label[:, :2] = (label[:, :2] / second_per_frame).astype(int)
label[:, 2] = label[:, 2] + 1
if len(label) > self.padded_len:
print(
"Warning: length of {} exceeds padded length".format(uttname))
label = label[:self.padded_len, :]
label_padded = np.zeros((self.padded_len, 3))
label_padded[:len(label), :] = label
return uttname, feat, label_padded, len(label)
def __getitem__(self, idx):
uttname = self.uttlist[idx]
data, samplerate = kaldi_data.load_wav(self.utt2ark[uttname])
Y = feature.stft(data, self.frame_size, self.frame_shift)
feat = feature.transform(Y, self.input_transform)
seg_list = self.utt2seg[uttname]
label = self.process_label(seg_list)
second_per_frame = self.frame_shift * 1.0 / self.rate
label[:, :2] = (label[:, :2] / second_per_frame).astype(int)
label[:, 2] = label[:, 2] + 1
return uttname, feat, label