sphinxtrain decode vs pocketspphinx_contionus for An4

Speech Recognition Toolkit

Brought to you by: air, arthchan2003, awb, bhiksha, and 5 others

This project can now be found here.

sphinxtrain decode vs pocketspphinx_contionus for An4

Forum: Help

Creator: cricket

Created: 2017-01-09

Updated: 2017-01-09

cricket - 2017-01-09

I followed the steps http://cmusphinx.sourceforge.net/wiki/tutorialam to train the an4 database(raw little endian)

testing using sphinxtrain -s decode run gave me WER of 18.5%

I ran pocketsphinx_continuous on the an4 test files using the models generated .

There is some discrepancy when I compare the pocketsphinx_continuous results with an4.align file results

For a raw test file with transcription "ONE THREE SEVEN" the an4.align shows correct results where as pocketsphinx_continuous shows "ONE THREE SECOND"

sphinxtrain decode and pocketspphinx_contionus both use the same hmm model ,dic and lm but NOT ABLE TO FIGURE why the difference in prediciton.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
- Nickolay V. Shmyrev - 2017-01-09
  
  Duplicate of https://sourceforge.net/p/cmusphinx/discussion/sphinx4/thread/989f13c4/
  
  If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

cricket - 2017-01-09

Thanks for the reply.
1. I set $CFG_CMN='none' in the sphinx_train.cfg file .The accuracy comes down but similiar result in sphinxtrain decode and pocketspphinx_contionus
2. With $CFG_CMN='batch' mode with cmninit values in the range 1 to 100 results did not match for sphinxtrain decode and pocketspphinx_contionus.
3. The other thing I got to know from the forum is using ps_process_raw .I will have to try that to run the pocketspphinx_contionus in batch mode

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

Log in to post a comment.