Menu

Correct options to send sphinx's mfc file into LIUM

Help
Emil Lundh
2016-04-26
2016-04-28
  • Emil Lundh

    Emil Lundh - 2016-04-26

    I have an .mfc file encoded by sphinx_fe. Now next step, I want to segment the audio with LIUM SpkDiarization, using this .mfc file as input.
    Trying LIUM options -fInputDesc sphinx doesn't quite help, because I receive the following complaint:

    08:01.088 MDecode        FINE  |     decoder.accumulation starting at 0 to 895  {make() / 10}
    08:01.094 Diarization    SEVERE| Exception error    {run() / 10}
    java.lang.ArrayIndexOutOfBoundsException: -1
    

    after a few quantities have evauated to NaN. Full output here: http://pastebin.com/Tp1EK9EZ
    I speculate that I have provided the wrong parameters. The parameter list is long (default seems to be audio2sphinx,1:1:0:0:0:0,13,0:0:0:0) and I have no clue what all these should be for a feature file from sphinx_fe. Does anybody know?

     
  • Emil Lundh

    Emil Lundh - 2016-04-26

    Sorry for the confusion, indeed I have set --fInputDesc sphinx. (See link above for the exact command line I used.) But there are 11 other parameters to set there. I have no clue here. It says they corresponds to energy, delta, umber of features, ... Any idea which ones are important? Because if I use default for them I obtain the exception cited above.

     

    Last edit: Emil Lundh 2016-04-26
  • Emil Lundh

    Emil Lundh - 2016-04-27

    To clarify: Setting --fInputDesc sphinx does NOT help; not alone.

    Probably one has to change some more settings. The question is which.

     
    • Nickolay V. Shmyrev

      hm, I think the right option should be not simply sphinx but something like

      sphinx,1:1:0:0:0:0,13,0:0:0:0
      

      Overall, I'm quite unhappy with LIUM, you'd better start writing code from scratch instead of looking on it. It is extremely hard to configure LIUM properly.

       
  • Emil Lundh

    Emil Lundh - 2016-04-28

    Thanks Nickolay, I guess I'll just experiment along. After all, I'm not really up to writing my own speech segmentation tool! (Were there plans for writing a segmentation library within CMU Sphinx?)
    I could also redesign my workflow to give LIUM a wav file, but that feels suboptimal for a number of (internal) reasons...

     

Log in to post a comment.