Menu

Pocketsphinx slow for grammar, fast for LM

Help
2012-06-11
2012-09-22
  • Nathan Glenn

    Nathan Glenn - 2012-06-11

    I made a decent grammar for recognizing numbers in pocketsphinx.
    Unfortunately, it's intolerably slow. Recognition is slow, and the application
    (PocketSphinxAndroidDemo) takes a very long time to start up. To get around
    that for now, I used the JSGF functions in Sphinx 4 to generate 200,000
    sentences and then used cmuclmtk to make a language model. PocketSphinx
    performs recognition very quickly with a language model.

    Are there any command line flags or other options that I can use to make
    pocketsphinx perform recognition quickly with a grammar? I am already using
    -backtrace false and -bestpath false.

     
  • Joseph S. Wisniewski

    Sphinx JSGF support is severely broken.

    If your grammar is a simple network of numbers, draw it as a network, number
    the nodes, and hand edit an FSG file. It will run fast and accurate.

     
  • Nickolay V. Shmyrev

    Sphinx JSGF support is severely broken.

    Please report about the issues you have found. So far we are not aware about
    "broken" JSGF.

     
  • Nathan Glenn

    Nathan Glenn - 2012-07-20

    Hmm... I don't think that hand editing will be an option. I did the numbers
    grammar to test if PocketSphinx would be slow with a small grammar, which it
    was. For now I am using a generated LM, but any ideas on this issue will be
    much appreciated.

     

Log in to post a comment.