I am working on speech recognition. I have used pocketsphinx tool for this. In Pocketsphinx tool there are two command are available for this - pocketsphinx_batch and pocketsphinx_continuous. I have notice that both command gives different result although I have used same dictionary, language model and hmm. The accuracy of pocketsphinx_batch is around 83% where as for pocketsphinx_continuous it is only 45%. Can any one tell me why such a large difference in accuracy of both command?
For recognition, I have used stored wav files. I am using pocketsphinx-5prealpha on ubuntu 14.04.
Any help in this regard will be appreciated.
Regards,
Shashikant
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Pocketsphinx_continuous analyses audio sequentially frame by frame. Batch analyses audio as a whole and normalizes volume properly. If you restart continuous on short samples it might be inaccurate.
Alternatively you can set channel parameters with -cmninit option of pocketsphinx_continuous, then it will be same as batch. You can set cmninit values to the ones that printed in batch logs.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi all,
I am working on speech recognition. I have used pocketsphinx tool for this. In Pocketsphinx tool there are two command are available for this -
pocketsphinx_batchandpocketsphinx_continuous. I have notice that both command gives different result although I have used same dictionary, language model and hmm. The accuracy ofpocketsphinx_batchis around 83% where as forpocketsphinx_continuousit is only 45%. Can any one tell me why such a large difference in accuracy of both command?For recognition, I have used stored wav files. I am using pocketsphinx-5prealpha on ubuntu 14.04.
Any help in this regard will be appreciated.
Regards,
Shashikant
Pocketsphinx_continuous analyses audio sequentially frame by frame. Batch analyses audio as a whole and normalizes volume properly. If you restart continuous on short samples it might be inaccurate.
Alternatively you can set channel parameters with -cmninit option of pocketsphinx_continuous, then it will be same as batch. You can set cmninit values to the ones that printed in batch logs.
Same discussion on github
https://github.com/cmusphinx/pocketsphinx/issues/116