Speech Recognition Engines

Poor accuracy with pocketsphinx
User: arun
Date: 8/19/2009 12:50 am
Views: 8421
Rating: 5


I am trying to make a simple dictation system using pocketsphinx. Following are the components I am using for making the Speech Recognition system

1) Pocketsphinx - as decoder

2) wsj_all_sc.cd_semi_5000 - as acoustic model

3) I used 'CMU-Cam_Toolkit_v2' and 'lm3g2dmp' tools for making the language model with 28K words dictionary. I used "SimpleLM.pl" tool present in 'CMU-Cam_Toolkit_v2' to derive the 28K dictionary from cmudict.0.7a_SPHINX_40 dictionary.

With this I am trying to convert a simpe 5 seconds recording into text. But I am getting a very low accuracy which is nearly 10%.

I have checked into the dictionary and the words in the recording are also present in the dictionary and the langauge model is also made using that dictionary as well. So what may be the problem due to which the accuracy is so bad.

Please help me out with any suggestions, pointers.

Do I need to change any models or dictionary or is it required to make a new acoustic model?

Its necessary for me to use pocketsphinx so I cannot change the decoder.

I would really appreciate any help/ suggestions.

Thanks a lot.


--- (Edited on 8/19/2009 12:50 am [GMT-0500] by ) ---

Re: Poor accuracy with pocketsphinx
User: nsh
Date: 8/19/2009 5:37 pm
Views: 5796
Rating: 6

Heh, I wrote a little bit about this issue today


--- (Edited on 8/19/2009 5:37 pm [GMT-0500] by nsh) ---