Reformer Model for Speech Recognition

[stefan-falk]

Coming from tensor2tensor I was wondering whether the Reformer model would be also a candidate for speech recognition? Looking at the examples there is none for ASR.

Would it be possible to train an ASR model on the Reformer or would code changes be necessary? If so, can we estimate how much would have to be changed on the model implementation?

Thank you for any insight into this!

[lukaszkaiser]

I believe Reformer (esp. with SRU as feed-forward, which is an hparam already) should make a nice ASR model. Didn't have time to work on it yet, but it'd be great to try!

[lukaszkaiser]

I don't think many changes are needed in terms of the model, but the input pipeline may need some thought. I believe that just feeding bytes could work, but it needs experimentation to see...

[stefan-falk]

It would be interesting to see. I guess translation (or text2text problems in general) could be tested as well?

[RegaliaXYZ]

@stefan-falk Have you managed to do the ASR problem?

[stefan-falk]

@RegaliaXYZ I didn't try it yet, sorry.