Re: [kaldi-help] WFST decoder for CTC models

[Daniel Povey]

I don't understand the issue about `<eps>` and the map.  The decoder already treats `<eps>` specially.

I would recommend to have the neural-net output indexes (plus one) as the ilabels in the 

decoding graph, and use latgen-faster.

However I just noticed that latgen-faster does not actually exist.

You could create it though, it would have similar differences as decode-faster-mapped vs. decode-faster.

On Tue, Jan 7, 2020 at 9:34 AM Shahram Ghorbani <ghorbani...@gmail.com> wrote:

Hi,

I have a CTC trained model and also TLG.fst for the decoding. But, since I have <eps> in my tokens I a map from my token to the model's output.  I was wondering if there is any way to map the tokens to the CTC model's output. I have seen an option like --tid-to-pdf=tid2pdf  for latgen-faster-mapped, but could not find this option anymore :) 

Thanks

Shahram

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/102b3b09-7a11-4eac-9bba-46f18dd5a8b1%40googlegroups.com.

[Shahram Ghorbani]

Thanks, Dan. 

Will try to edit the code. 

On Monday, January 6, 2020 at 8:49:33 PM UTC-6, Dan Povey wrote:

I don't understand the issue about `<eps>` and the map.  The decoder already treats `<eps>` specially.

I would recommend to have the neural-net output indexes (plus one) as the ilabels in the 

decoding graph, and use latgen-faster.

However I just noticed that latgen-faster does not actually exist.

You could create it though, it would have similar differences as decode-faster-mapped vs. decode-faster.

On Tue, Jan 7, 2020 at 9:34 AM Shahram Ghorbani <ghorban...@gmail.com> wrote:

Hi,

I have a CTC trained model and also TLG.fst for the decoding. But, since I have <eps> in my tokens I a map from my token to the model's output.  I was wondering if there is any way to map the tokens to the CTC model's output. I have seen an option like --tid-to-pdf=tid2pdf  for latgen-faster-mapped, but could not find this option anymore :) 

Thanks

Shahram

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi...@googlegroups.com.