Re: kaldi-DNN

[Daniel Povey]

I recommend to start clean with a pytorch-based approach e.g. based on ctc; that performs better these days with modern neural net architectures.  It should be quite easy to find something.  

On Mon, Jan 19, 2026 at 6:27 PM Hema A. Murthy <he...@cse.iitm.ac.in> wrote:

Dear Dan Povey

       We are trying to replace the emission probabilities obtained from DNN with our own custom probabilities.

Could somebody help us with the files that are relevant?  There are way too many NN codes in kaldi-tree.

All of us are pretty new to kaldi itself.
Thanks, Hema A Murthy

[Hema A. Murthy]

Thanks.  Alignments using CTC are not as good as HMM+Signal Processing for Indian language data.

-hema

[Daniel Povey]

oh i see.  for alignment you could probably use one of the standard GMM recipes.

--
Go to http://kaldi-asr.org/forums.html to find out how to join the kaldi-help group
---
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 visit https://groups.google.com/d/msgid/kaldi-help/3f550166-1527-4701-9a21-2f1fe1819ff8n%40googlegroups.com.

[ASR_OCEAN]

Hi Dan!

I have also a related issue.

I am trying to find phoneme alignment using pytorch CTC. 

For TIMIT dataset it works very well (training loss goes down below 1.0) 

but for non-English language such as Evenki language (https://doreco.huma-num.fr/preview/languages/even1259), 

the CTC loss is between 2 and 3 and the predictions is not correct.

Thank you.

[Daniel Povey]

maybe you have not enough data to train that language.  you could maybe try fine-tuning a system already trained for a previous language, that might work better.

but GMM-based recipes are definitely more stable.

To view this discussion visit https://groups.google.com/d/msgid/kaldi-help/9ef83d30-51dd-4c2d-8dba-ea909528ab70n%40googlegroups.com.