How to use my trained GMM acoustic model along with a decoding graph?

[lucife...@gmail.com]

I have trained a GMM acoustic model along with a decoding graph , how can I start using it?

I have followed the blog by Josh Meyer's -http://jrmeyer.github.io/asr/2016/09/12/Using-built-GMM-model-Kaldi.html

but i ended up creating an empty transcript 

My model was trained at wer =16.50

this steps show error at the time of generating lattice

/home/lucifer/kaldi/src/gmmbin/gmm-latgen-faster --max-active=14000 --beam=12.0 --lattice-beam=6.0 --acoustic-scale=0.083333 --allow-partial=true --word-symbol-table=/home/lucifer/kaldi/testing/words.txt \
> /home/lucifer/kaldi/egs/digits/exp/tri1/final.mdl /home/lucifer/kaldi/egs/digits/exp/tri1/graph/HCLG.fst ark:/home/lucifer/kaldi/egs/digits/transcriptions/delta-feats.ark \
>    ark,t:/home/lucifer/kaldi/new_test/decode/lattices.ark
/home/lucifer/kaldi/src/gmmbin/gmm-latgen-faster --max-active=14000 --beam=12.0 --lattice-beam=6.0 --acoustic-scale=0.083333 --allow-partial=true --word-symbol-table=/home/lucifer/kaldi/testing/words.txt /home/lucifer/kaldi/egs/digits/exp/tri1/final.mdl /home/lucifer/kaldi/egs/digits/exp/tri1/graph/HCLG.fst ark:/home/lucifer/kaldi/egs/digits/transcriptions/delta-feats.ark ark,t:/home/lucifer/kaldi/new_test/decode/lattices.ark
WARNING (gmm-latgen-faster[5.4.269~1-66145]:ProcessNonemitting():lattice-faster-decoder.cc:819) Error, no surviving tokens: frame is -1
Akash_0.wav
LOG (gmm-latgen-faster[5.4.269~1-66145]:DecodeUtteranceLatticeFaster():decoder-wrappers.cc:286) Log-like per frame for utterance Akash_0.wav is -10.3948 over 4084 frames.
LOG (gmm-latgen-faster[5.4.269~1-66145]:main():gmm-latgen-faster.cc:176) Time taken 1.74808s: real-time factor assuming 100 frames/sec is 0.0428032
LOG (gmm-latgen-faster[5.4.269~1-66145]:main():gmm-latgen-faster.cc:179) Done 1 utterances, failed for 0
LOG (gmm-latgen-faster[5.4.269~1-66145]:main():gmm-latgen-faster.cc:181) Overall log-likelihood per frame is -10.3948 over 4084 frames

Thanks in advance

---

How to use an Existing GMM Recognizer for Decoding in Kaldi

[Daniel Povey]

It looks to me like your decoding graph (HCLG.fst) may have a problem.
I don't have time to debug your setup, especially since you were
following an external tutorial. I suggest to use resources like the
HTK Bool (to understand the basic of ASR) and the Kaldi for Dummies
tutorial and the like, for a while.

Dan

> --
> 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/adae76bc-d57d-4b09-a3cf-7e5bf884f2d8%40googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[lucife...@gmail.com]

I have followed Kaldi for Dummies tutorial and generated hclg.fst

after following Kaldi for Dummies tutorial I got WER 16.50 for mono and 14.50 for triphone training.

My problem is to use this trained model

[Daniel Povey]

I am hoping someone else can respond to this, the answer is not that brief.  Also it may depend what you mean by "use" the model, i.e. what your application is.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/23cd84f3-97e3-4687-a187-26b9b534ef84%40googlegroups.com.

[lucife...@gmail.com]

I want to use my trained model to generate transcript on sample audio file , Is there any official blog ?

[Shin XXX]

/home/lucifer/kaldi/src/gmmbin/gmm-latgen-faster --max-active=14000 --beam=12.0 --lattice-beam=6.0 --acoustic-scale=0.083333 --allow-partial=true --word-symbol-table=/home/lucifer/kaldi/testing/words.txt \ 

 

Looks like you're looking for some commands to directly generate the transcripts.

run the default decode.sh, for chain model it should be steps/nnet3/decode.sh, you'll find the command you need in log files(decode.*.log). The default log directory is under the decode directory. 

Shin

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/ca4bf107-1b5e-434a-a908-b9fddfaa3c5d%40googlegroups.com.

[Daniel Povey]

The tricky thing is that you need to have previously generated the MFCC and CMVN stats.  The easiest thing, if you want to do this on a bunch of data (not just one file) is to set it up as a 'data' directory as described in the 'data preparation' section of the Kaldi docs-- just without the 'text' file, if you don't have that.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CADtO4NV2BvUrTYiDqzkk5ePPY1TYMkDeAFYRJKzdaoZySzQpvA%40mail.gmail.com.