def read_recognizer(inference_config_or_name='latest'):
# download specified model automatically if no model exists
if len(get_all_models()) == 0:
download_model('latest')
# create default config if input is the model's name
if isinstance(inference_config_or_name, str):
model_name = resolve_model_name(inference_config_or_name)
inference_config = Namespace(model=model_name, device_id=-1, lang='ipa', approximate=False)
else:
assert isinstance(inference_config_or_name, Namespace)
inference_config = inference_config_or_name
model_path = Path(__file__).parent / 'pretrained' / inference_config.model
if inference_config.model == 'latest' and not model_path.exists():
download_model(inference_config)
assert model_path.exists(), f"{inference_config.model} is not a valid model"
# create pm (pm stands for preprocess model: audio -> feature etc..)
pm = read_pm(model_path, inference_config)
# create am (acoustic model: feature -> logits )
am = read_am(model_path, inference_config)
# create lm (language model: logits -> phone)
lm = read_lm(model_path, inference_config)
return Recognizer(pm, am, lm, inference_config)
def read_recognizer(inference_config):
model_path = Path(__file__).parent / 'pretrained' / inference_config.model
if inference_config.model == 'latest' and not model_path.exists():
download_model(inference_config)
assert model_path.exists(
), f"{inference_config.model} is not a valid model"
# create pm (pm stands for preprocess model: audio -> feature etc..)
pm = read_pm(model_path, inference_config)
# create am (acoustic model: feature -> logits )
am = read_am(model_path, inference_config)
# create lm (language model: logits -> phone)
lm = read_lm(model_path, inference_config)
return Recognizer(pm, am, lm, inference_config)
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()