pocketsphinx decoding with dnn models

[bhargav]

Hi everyone

I am trying to build DNN based speech recognition system. I am able to code the training process of DNNs. However, I am only half-successful in building ASR because I am not able to write the decoder part.
Hence I would like to know if it is possible to incorporate likelihoods, obtained from DNN, in pocketsphinx decoder.

Any help is appreciated

Thanks
Bhargav

[Arseniy Gorin]

Decoding from phone likelihoods/posteriors is not implemented in pocketsphinx. You will need to re-write quite a lot of code, but yes, in theory there is no problem in doing that.

You better look at kaldi project - DNNs are already implemented there.

[Nickolay V. Shmyrev]

Actually reading from senone score file is implemented. ps_decode_senscr function does that. You just need to prepare a file in proper format.

[Arseniy Gorin]

nice to know, thanks

[bhargav]

Thanks Arseniy and Nickolay

I was going through the code. I came across this quantity named acmod->senone_scores. Is this the one we can play around with?

[bhargav]

HI Nickolay
As you mentioned, I saw that ps_decode_senscr is being used in pocketsphinx_batch with -senin argument. I am trying to run following command

pocketsphinx_batch -cepdir feat -ctl batch_file -dict my_dict.dic -lm my.lm -senin senone_score -hmm my_models
where senone_score is a text file.

The above command is giving following error
ERROR: "cmd_ln.c", line 627: Bad argument value for -senin: senone_score
Kindly, let me know if text file can be used or senone score dump file is a another format

Thanks
Bhargav

[Nickolay V. Shmyrev]

Kindly, let me know if text file can be used or senone score dump file is a another format

The options must be -senin yes -cepdir . -cepext dump. Dumps for each utterance must be in a cepdir folder and should have dump extension then.

-senin works like -adcin.