def recognize(self, filename, lang_id):
# recognize a single file
assert str(filename).endswith(
'.wav'), "only wave file is supported in allosaurus"
# load wav audio
audio = read_audio(filename)
# extract feature
feat = self.pm.compute(audio)
# add batch dim
feats = np.expand_dims(feat, 0)
feat_len = np.array([feat.shape[0]], dtype=np.int32)
tensor_batch_feat, tensor_batch_feat_len = move_to_tensor(
[feats, feat_len], self.config.device_id)
tensor_batch_lprobs = self.am(tensor_batch_feat, tensor_batch_feat_len)
if self.config.device_id >= 0:
batch_lprobs = tensor_batch_lprobs.cpu().detach().numpy()
else:
batch_lprobs = tensor_batch_lprobs.detach().numpy()
token = self.lm.compute(batch_lprobs[0], lang_id)
return token
def recognize(self, filename, lang_id='ipa', topk=1, emit=1.0, timestamp=False):
# recognize a single file
# filename check (skipping for BytesIO objects)
if not isinstance(filename, BytesIO):
assert str(filename).endswith('.wav'), "only wave file is supported in allosaurus"
# load wav audio
audio = read_audio(filename)
# extract feature
feat = self.pm.compute(audio)
# add batch dim
feats = np.expand_dims(feat, 0)
feat_len = np.array([feat.shape[0]], dtype=np.int32)
tensor_batch_feat, tensor_batch_feat_len = move_to_tensor([feats, feat_len], self.config.device_id)
tensor_batch_lprobs = self.am(tensor_batch_feat, tensor_batch_feat_len)
if self.config.device_id >= 0:
batch_lprobs = tensor_batch_lprobs.cpu().detach().numpy()
else:
batch_lprobs = tensor_batch_lprobs.detach().numpy()
token = self.lm.compute(batch_lprobs[0], lang_id, topk, emit=emit, timestamp=timestamp)
return token
def prepare_feature(data_path, model):
model_path = Path(__file__).parent.parent / 'pretrained' / model
# create pm (pm stands for preprocess model: audio -> feature etc..)
pm = read_pm(model_path, None)
# data path should be pointing the absolute path
data_path = data_path.absolute()
# writer for feats
feat_writer = KaldiWriter(data_path / 'feat')
# writer for the shape of each utterance
# format: utt_id shape[0] shape[1]
shape_writer = open(data_path / 'shape', 'w')
for line in tqdm(
open(data_path / 'wave', 'r', encoding='utf-8').readlines()):
fields = line.strip().split()
utt_id = fields[0]
audio_path = fields[1]
assert Path(audio_path).exists(), audio_path + " does not exist!"
audio = read_audio(audio_path)
# extract feature
feat = pm.compute(audio)
# write shape
shape_writer.write(f'{utt_id} {feat.shape[0]} {feat.shape[1]}\n')
feat_writer.write(utt_id, feat)
feat_writer.close()
shape_writer.close()
def recognize(self, filename, lang_id):
# load wav audio
audio = read_audio(filename)
# extract feature
feat = self.pm.compute(audio)
# add batch dim
feats = np.expand_dims(feat, 0)
feat_len = np.array([feat.shape[0]], dtype=np.int32)
tensor_batch_feat, tensor_batch_feat_len = move_to_tensor(
[feats, feat_len], self.config.device_id)
tensor_batch_lprobs = self.am(tensor_batch_feat, tensor_batch_feat_len)
if self.config.device_id >= 0:
batch_lprobs = tensor_batch_lprobs.cpu().detach().numpy()
else:
batch_lprobs = tensor_batch_lprobs.detach().numpy()
token = self.lm.compute(batch_lprobs[0], lang_id)
return token