Decode phone level without language model

[miamoto9]

Hey!

 I have a chain model and want to perform decoding for single utterances without LM at the phone level only. My goal is to see how the model behaves at an acoustic level only. 

Is there an easy way of performing this operation in Kaldi? 

Thanks a lot,

Carlos

[nick....@avinium.com]

If I've understood you correctly, you can achieve this by creating a phone-only language model, loosely:

1) Create lexicon.txt and words.txt that maps phones->phones

b b

ah ah

k k

(etc etc)

2) Use prepare_lang.sh with --position-dependent-phones=false to create L.fst

3) Create a 1-gram .arpa file representing a language model that maps phones->phones and compile with arpa2fst

4) Call mkgraph.sh with these files to create HCLG.fst and pass that to your decoder

These might not be the precise filenames/arguments but roughly that's the gist. Other people have asked similar questions on this list before so you can search around a bit if you need more hints. Alternatively there is a binary that you can use to get the whole phone lattice but this will be after language model decoding/scoring (I believe - someone correct me if I'm wrong) so this isn't exactly what you're after.

[Amol Bole]

Same thing I have done. You can see the below thread. I have successfully built the graph as well but while decoding simply it is printing code.

For more information I am attaching my script here.

https://groups.google.com/g/kaldi-help/c/DSvr5_26LqQ

[nick....@avinium.com]

Are you asking for help? I assume "2.wav NG" is not the decoding result you're looking for, you've probably gone wrong somewhere in compiling the graph. You'll probably need to post the files (or at least the commands you used to produce them) to get meaningful help.

[Amol Bole]

It will good if you help me. Let me what I have to share?