Menu

Sphinxtrain -s decode run Vs Pocketsphinx_continuous -infile - discrepancies in results - how to get consistent results

Help
Q3Varnam
2018-04-28
2018-05-04
  • Q3Varnam

    Q3Varnam - 2018-04-28

    I have my own acoustic model. My language model is a jsgf file. I have about 200 test transcription recordings. After sphinxtrain creates the acoustic model, when it runs the decode stage and also when I run decode stage independently using sphinxtrain, I get a 96% success rate with these 200 test transcriptions/recordings.

    When I switch to pocketsphinx_continuous and use infile option and point to one of the recordings from the 200 odd recordings, it invariable has an error - i.e the same file, which was successful in sphinxtrain decode comes with a word error, I have tried it on all the 200 recordings and all of them have some error when run in pocketsphinx_continuous mode.

    I have tried using the beam width and other parameters that are set in the sphinx_train.cfg file to the command line.

    pocketsphinx_continuous -hmm ./model_parameters/tamilbus.ci_cont -jsgf ./etc/tamilbus.jsgf -infile ./wav/kks_1/kks_ticket_01.wav -dict ./etc/tamilbus.dic -beam 1e-80 -wbeam 1e-40 -lw 10 -wip 0.2

    Am I missing some configuration in pocketsphinx_continuous mode which is resulting in this error?

    Edit: When I mean error - I mean incorrect words being recognised than what is actually spoken.

     

    Last edit: Q3Varnam 2018-04-28
  • Q3Varnam

    Q3Varnam - 2018-04-28

    I have fixed this problem by adding cmninit 60,3,1 in addition to the beam values. Now pocket_sphinx_continuous is recognising the test files with the same accuracy as that of the sphinx train decode mode.

     

Log in to post a comment.

Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.