LM fusion with for ctc model

[rwth-i6]

LM fusion with for ctc model

[albertz]

Multiple things:

• E.g. CtcLoss and related code just wraps whatever TF provides. TF also provides ctc_beam_search_decoder as you mention, but that is without LM fusion. So far that I know, TF does not support CTC beam search with a LM. But as you mention, I have seem some code elsewhere which basically adds this to TF. I think you might find it in Mozilla DeepSpeech. This probably is the simplest way, to adapt this.
• We usually use our RASR software to perform decoding with hybrid HMM/NN models. CTC is just a special case of a HMM, so you can use RASR to do the CTC beam search, also with a language model. But getting this to work is non-trivial, and requires quite some effort.

[albertz]

Not sure if there is an issue here on this repository, so I'm closing this. If you think there is some bug in RETURNN, please report for that repository.