def feature_encoders(self, _):
return {
"inputs": None, # Put None to make sure that the logic in
# decoding.py doesn't try to convert the floats
# into text...
"waveforms": audio_encoder.AudioEncoder(),
"targets": ByteTextEncoderWithEos(),
}
def feature_encoders(self, data_dir):
if self.is_8k:
sample_rate = 8000
else:
sample_rate = 16000
encoders = {
"inputs": None,
"waveforms": audio_encoder.AudioEncoder(sample_rate=sample_rate)
}
vocab_file = os.path.join(data_dir, "vocab",
"bpe.%d.t2t" % self.approx_vocab_size)
target_encoder = SubwordTextEncoderWithEos(filename=vocab_file)
encoders["targets"] = target_encoder
return encoders