How can I get phoneme-level and word-level segmentation during online recognition

[Zibo Meng]

Dear all, 

I am now using fisher_english pretrained nnet2 model to do online decoding. Can I get the phoneme level segmentation?

Thanks

[Daniel Povey]

You can get it the same way you would using any decoder, from the lattice.

You can use lattice-best-path (with the appropriate acoustic scale) and  pipe the 'alignments-wspecifier' into ali-to-phones.  ali-to-phones has various options to control the formatting.

Dan

--
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.
For more options, visit https://groups.google.com/d/optout.

[Zibo Meng]

Dan,

Thank you very much!

It really helped.

Best,

Zibo

[Zibo Meng]

Dear Dan,

I have a follow up question.

I tried to use the pretrained nnet2-online model trained on librispeech from http://kaldi-asr.org/downloads/build/6/trunk/egs/librispeech/s5/exp/nnet2_online/, and get the graph from http://kaldi-asr.org/downloads/build/6/trunk/egs/librispeech/s5/exp/tri5b/graph_tgsmall/, using the following command:

online2-wav-nnet2-latgen-faster --do-endpointing=false     --online=false     --config=nnet_a_online/conf/online_nnet2_decoding.conf     --max-active=7000 --beam=15.0 --lattice-beam=6.0     --acoustic-scale=0.1 --word-symbol-table=graph/words.txt    nnet_a_online/final.mdl graph/HCLG.fst "ark:echo utterance-id1 utterance-id1|" "scp:echo utterance-id1 cut.wav|"    ark:/dev/null

However, I only get the result : utterance-id1 <UNK> <UNK>

Did do something wrong?

Thanks,

Best,

Zibo

On Wednesday, November 25, 2015 at 4:46:15 PM UTC-5, Dan Povey wrote:

[Daniel Povey]

If you set the verbose level higher you can see the objective function improvement from iVector estimation.  If it is very large (more than 10 or 20 or so) it generally indicates severe acoustic-environment mismatch.  This can sometimes lead to very bad decoding results.  There  is a configuration variable to that program that can be used to downweight the silence in the iVector estimation; if it is set to something like 0.001 and you also supply the list of silence phones from lang/phones/silence.csl (or something like that), it can sometimes help improve robustness.

Dan

[Zibo Meng]

Dear Dan,

Thank you very much!

Would you please give the name of the configuration variable you mentioned?

Best,

Zibo