search

awni / speech

A PyTorch Implementation of End-to-End Models for Speech-to-Text
Apache License 2.0
747 stars 175 forks source link

RNN Transducer inference problem #26

Closed HawkAaron closed 6 years ago

HawkAaron commented 6 years ago

Hi, In your transducer_model inference function: https://github.com/awni/speech/blob/master/speech/models/transducer_model.py#L93 where use both acoustic feature and ground truth label, which is not inference at all. At least, do gready search or beam search.

When training model using transducer loss, the acoustic model PER is too big, can you provide a trained baseline of RNN Transducer ?

awni commented 6 years ago

Yes it still uses the ground truth previous label, this needs to be changed. Marking this as a duplicate of #21