Menu

Communicator acoustic model gives SegFault

Help
mindthief
2010-03-04
2012-09-22
  • mindthief

    mindthief - 2010-03-04

    Dear all,
    I created the Pocketsphinx "hello world" application (from http://cmusphinx.s
    ourceforge.net/wiki/PocketSphinxMigration).
    When I use the Communicator acoustic model, it
    gives me this error message:

    ERROR: "fe_interface.c", line 100: FFT: Number of points must be greater or
    equal to frame size (409 samples)
    Segmentation fault

    It works fine when I use the acoustic models wsj1, wsj16k, wsj8k, hub4 (4000),
    and hub4 (6000), but I am interested in using Communicator in my application.
    I got all the models from http://www.speech.cs.cmu.edu/sphinx/models/
    Advice will be appreciated. Thanks!

     
  • Nickolay V. Shmyrev

    You probably want to use -samprate 8000

     
  • mindthief

    mindthief - 2010-03-04

    hmm that did help! but not sure if it's working correctly yet. I made the
    following addition:

    config = cmd_ln_init(NULL, ps_args(), TRUE,
    "-samprate", "8000",
    "-hmm", MODELDIR "/am/communicator",
    "-lm", MODELDIR "/lm/turtle/turtle.lm",
    "-dict", MODELDIR "/lm/turtle/turtle.dic",
    NULL);

    ...
    (is this right?) and now it doesn't give me a SegFault anymore. But at the end
    when it should say "Recognized: GO FORWARD TEN METERS"
    it instead says "Recognized: " (it is blank)

    Is this expected behavior? Or is something still wrong?

     
  • Nickolay V. Shmyrev

    Is this expected behavior? Or is something still wrong?

    The input audio must be at 8kHz, isn't it? Communicator is the model for
    telephone speech.

     
  • mindthief

    mindthief - 2010-03-04

    Yeah I think the input audio I'm using (provided by PocketSphinx) is 16KHz.. I
    assumed that since it is HIGHER quality, that if communicator samples it at
    8KHz it would still get a good sample.... but I guess this isn't how it works?
    Does the input audio always have to be sampled at the same freq as the
    acoustic model is trained on?

     
  • Nickolay V. Shmyrev

    Does the input audio always have to be sampled at the same freq as the
    acoustic model is trained on?

    Right now it must be the same frequency.

     
  • mindthief

    mindthief - 2010-03-04

    okay, thanks a lot!

     
  • Mariano Cerdeiro

    I had the same issue and I have fixed using the parameter -mftt=512

    Regards,
    Mariano.-

     
  • Nickolay V. Shmyrev

    -mftt=512

    It should be

    -nfft 512
    

    without =. The proper solution would be to remove -nfft 256 from feat.params
    file inside the model.

     

Log in to post a comment.