def _load_feature(batch):
""" Load acoustic feature from files.
The features have been prepared in previous step, usualy by Kaldi.
Args:
batch: a list of tuple (wav id , feature ark path).
Returns:
(keys, feats, labels)
"""
keys = []
feats = []
lengths = []
for i, x in enumerate(batch):
try:
mat = kaldi_io.read_mat(x[1])
feats.append(mat)
keys.append(x[0])
lengths.append(mat.shape[0])
except (Exception):
# logging.warn('read utterance {} error'.format(x[0]))
pass
# Sort it because sorting is required in pack/pad operation
order = np.argsort(lengths)[::-1]
sorted_keys = [keys[i] for i in order]
sorted_feats = [feats[i] for i in order]
labels = [x[2].split() for x in batch]
labels = [np.fromiter(map(int, x), dtype=np.int32) for x in labels]
sorted_labels = [labels[i] for i in order]
return sorted_keys, sorted_feats, sorted_labels
def _load_from_file(batch):
keys = []
feats = []
lengths = []
for i, x in enumerate(batch):
try:
mat = kaldi_io.read_mat(x[1])
feats.append(mat)
keys.append(x[0])
lengths.append(mat.shape[0])
except (Exception):
# logging.warn('read utterance {} error'.format(x[0]))
pass
# Sort it because sorting is required in pack/pad operation
order = np.argsort(lengths)[::-1]
sorted_keys = [keys[i] for i in order]
sorted_feats = [feats[i] for i in order]
labels = [x[2].split() for x in batch]
labels = [np.fromiter(map(int, x), dtype=np.int32) for x in labels]
sorted_labels = [labels[i] for i in order]
return sorted_keys, sorted_feats, sorted_labels
