Implementation of MoChA for ASR

[rwth-i6]

Implementation of MoChA for ASR

[albertz]

No, that config corresponds to the model we describe in the paper An analysis of local monotonic attention variants. It's much simpler than MoChA, but it is also online capable. We currently do not have created a config for MoChA.

[manish-kumar-garg]

Thanks @albertz for clarifying.
How can I use this model for online decoding? Is there any implementation already available?

[albertz]

No, not really. The encoder in that config is also still a BLSTM. You would have to change it to become a unidirectional LSTM, or some other online model (e.g. LC BLSTM, or chunked BLSTM, or unidir Transformer encoder, or CNN, ...), and you probably need some tuning first to get good performance. Then you would need to develop some decoding pipeline where you can stream new audio in, run the encoder on that new audio, and successively output new hypotheses. We are working on sth like that, but it's not ready and not published, and there is no timeline. But if you invest some work into it yourself, it might actually not be too complicated.

[manish-kumar-garg]

Got it. Thanks!