Menu

low accuracy on Sphinx3.5 with sample data

Help
2005-10-17
2012-09-22
  • Stan Stelle

    Stan Stelle - 2005-10-17

    I am just starting to investigate Sphinx3.5 and am trying to run some sample data through the decoder.exe to make sure my setup is correct before trying my own audio input.
    I have downloaded the open source sample acoustic and language models. I am inputting the file an391-mjwl-6.mfc from the AN4 database. According to the transcript it just contains
    the words ENTER, TWO, NINE, EIGHT and ONE. When I run the decoder on it, it gets 0% recognition. If I use the an4.ug.lm.DMP language model which I got from another link on the Sphinx web site, it only gets 2 out of 5 words, or 40% recognition. Below is the contents of the default.arg file I am feeding to the decoder. InputControl.txt just contains the single .MFC file. Am I specifying something incorrect? Am I missing a parameter? I was expecting a much higher recognition rate using the provided sample data.

    -mdef hub4opensrc.6000.mdef
    -mean means
    -var variances
    -mixw mixture_weights
    -tmat transition_matrices
    -subvq 8gau.6000sen.quant
    -dict cmudict.06d
    -fdict fillerdict.txt
    -lm language_model.arpaformat.DMP
    -hypseg hypseg.txt
    -ctl inputcontrol.txt
    -logfn log.txt

    Thanks

     
    • shiosai

      shiosai - 2005-10-18

      Hello,

      you can find some instructions and example files in sphinx3 sourcecodedir->model->lm->an4. The right parameters are very important for a good recognition so i would suggest to look into the args.an4* files for some good parameters.

      I hope this helps
      shio

      PS:
      you can look for the filesin the cvs, too:
      http://cvs.sourceforge.net/viewcvs.py/cmusphinx/sphinx3/model/lm/an4/

       

Log in to post a comment.