Mismatch in number of pdfs

[aliiire...@gmail.com]

Hi 

I trian models in tri1,tri2,tri3, sgmm2_4 methods, used FARSDAT egs.

and  decode another corpus by these models in tri2,tri3, sgmm2_4 but 

in the tri1 model its error:

decode.sh: feature type is delta 

Mismatch in number of pdfs with exp/tri1/final.mdl

what's wrong?

thanks

[Shin XXX]

Looks like the mismatch between graph and final.mdl. It would be helpful if you could attach the full decode log.

Shin

--
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 post to this group, send email to kaldi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/ecdd544d-d4ef-4bf5-b187-d92cf1cc7a02%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[aliiire...@gmail.com]

thanks

there is no decode log, only this error in terminal.

the number of pdf in my graph for my new corpus is 96.

and num_pdf of graph in my train(test set) corpus was 2056.

what can another information share?

how can fix it?

[Shin XXX]

Got it. You trained a model with 2056 pdfs, and you already got a HCLG.fst with 96 pdfs, then you tried to decode your new corpus with the 2056-pdf acoustic model and the 96-pdf HCLG.fst, therefore the mismatch happened. 

The HCLG.fst is the product of combining a language model(G.fst) and an acoustic model(final.mdl), if you want to decode sth with a new acoustic model, you should rebuild the HCLG.fst using ./utils/mkgraph.

Shin

--

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 post to this group, send email to kaldi...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/6839a41f-722e-4173-8d75-59b9a87fe022%40googlegroups.com.

[Shin XXX]

...sorry, just to clarify, HCLG is not a just simple combination of G.fst and final.mdl, the comments of ./utils/mkgraph.sh contains a better explanation. 

Shin

[aliiire...@gmail.com]

so thanks

i made graph but it's said

utils/mkgraph.sh: exp/test/tri1/graph/HCLG.fst is up to date.

so delete it and make it again, it successfully decode.

sorry, thanks