def read_and_preprocess_batched_audio(
self, ks,
exs):
"""Returns batched model input, audio, and sr."""
audios = []
for k, ex in zip(ks, exs):
audio = self._read_audio_and_resample(k, ex)
if audio.ndim > 1:
raise ValueError(f'Audio was too many dims: {audio.ndim}')
audios.append(audio)
sr = self._target_sample_rate
# Do some chunking.
if self._chunk_len:
logging.info('Chunk len: %s', self._chunk_len)
chunked_audios = []
for audio in audios:
if audio.shape[0] >= self._chunk_len:
chunk = utils.get_chunked_audio_fn(audio, self._chunk_len)
else:
chunk = np.expand_dims(audio, -1)
chunked_audios.append(chunk)
audios = np.concatenate(chunked_audios, axis=0)
audios = [audios[i] for i in range(audios.shape[0])]
# Convert audio to features, if required.
model_inputs = [self._audio_to_features(a, sr) for a in audios]
for model_input in model_inputs:
if model_input.shape != model_inputs[0].shape:
raise ValueError(f'Model input shapes not the same: {model_inputs}')
logging.info('model_input shape: %s', model_input.shape)
batched_model_input = np.stack(model_inputs, axis=0)
return batched_model_input, audios, sr
def tfex_to_chunked_audio(self, k, ex):
# Read audio from tf.Example, get the sample rate, resample if necessary,
# and convert to model inputs (if necessary).
model_input, sample_rate = self.read_and_preprocess_audio(k, ex)
# Do some chunking.
if self._chunk_len:
logging.info('Chunk len: %s', self._chunk_len)
if model_input.shape[0] >= self._chunk_len:
model_input = utils.get_chunked_audio_fn(
model_input, self._chunk_len)
logging.info('model_input after chunking: ')
return model_input, sample_rate