mkgraph_lookahead.sh stuck on fstdeterminizestar for HCL for specific words

[Mustafa Ebrar Aktaş]

I am trying to create lookahead graph for my already trained position independent kaldi model. First I followed the mkgraph_lookahead.sh and it stuck on composition of Ha.fst and CL_$N_$P.fst . 

There are similar issues below:

- https://groups.google.com/g/kaldi-help/c/pnXb_2ph4Ro/m/SwTX2kNgAwAJ

        It is suggested to use chain model, but I am using chain model already.

- https://github.com/kaldi-asr/kaldi/issues/4143

        It is suggested to omit fstdeterminizestar

Then I have tried to narrow down the issue. I observed that using some mini lexicons allows me to create HCLr.fst and script completes successfully.

Here are the different mini lexicons and their results:

• Words are "GO" and "I" (it has single phone pronunciation), can't complete

GO G OW

I AY

<UNK> SPN

• Words are "GO" and "IS", completes successfully.

GO G OW

IS IH Z

<UNK> SPN

• I tried a small hack by changing pronunciation of "I" like below, it completes successfully.

GO G OW

I AY AY

<UNK> SPN

I thought it was related to single phone pronunciations. However, the lexicon below makes it stuck, while it completes successfully if I remove anyone of the words.

AS AE Z

AS EH Z

AT AE T

FOR F AO R

FOR F ER0

FOR F R ER0

<UNK> SPN

I think it is related to transitions introduced by specific contexts, they make HCL nondeterminizable for my tree and model. Because even if I use single word lexicons with "I AY" or "I AY AY" , the latter one completes while the former fails. Because the former one has extra contexts below (you can find full lists in attached files):

SPN/AY/SIL

SPN/AY/SPN

SPN/AY/<eps>

SIL/AY/SIL

SIL/AY/SPN

SIL/AY/<eps>

<eps>/AY/SIL

<eps>/AY/SPN

<eps>/AY/<eps>

My question is that is it possible to make HCL determinizable? Because omitting fstdeterminizestar makes output HCLr.fst much bigger. (which has redundancy and slows down decoding)

I can share my model and tree if it is necessary.

[Daniel Povey]

Are you sure you are using a chain model, because those phone contexts you displayed are triphone, and all the examples of chain models that we have use left-biphone context.

--
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 on the web visit https://groups.google.com/d/msgid/kaldi-help/0c208e4d-cf84-4dba-90de-6422e2db9184n%40googlegroups.com.

[Mustafa Ebrar Aktaş]

Yes, it is a chain model. I have changed the model configuration for specific usage. Here is the output of tree-info:

num-pdfs 5872

context-width 3

central-position 1

Is the issue caused by using triphone context? 

[Daniel Povey]

Yes, that's likely the problem.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/cdf7255c-d3a7-409d-8c25-f9c6175bb7f0n%40googlegroups.com.

[Mustafa Ebrar Aktaş]

Using left-bigram seems like the solution, I could successfully created determinized HCLr with that. 

Thanks.