Menu

problem with sphinxtrain, bw

Help
2002-12-15
2012-09-22
  • Stephen Crowley

    Stephen Crowley - 2002-12-15

    I'm using the latest SphinxTrain CVS. I've setup my model according to tinydoc.txt, and get as far as the ./scripts_pl/02.ci_schmm/slave_convg.pl step.

    It runs for a bit, then I get the following error:

    Using 1 files: 0% ERROR:  12864 1941 ERROR: "backward.c", line 929: alpha(5.047836e-01) <> sum of alphas * betas (4.996917e-01) in frame 13415ERROR: ERROR: "baum_welch.c", line 279: teacher_in_space ignoredFinished

    The bw log is located <A HREF="http://intolerance.digitalpassage.com/misc/nasa.1-1.bw.log">here</A>

    What does this mean, or how could I possibly find out?

    Thanks,
    Stephen

     
    • Stephen Crowley

      Stephen Crowley - 2002-12-15

      Here is the correct url

      <A HREF="http://intolerance.digitalpassage.com/misc/nasa.1-1.bw.log">http://intolerance.digitalpassage.com/misc/nasa.1-1.bw.log</A>

       
    • Carl Quillen

      Carl Quillen - 2002-12-16

      It looks like the trainer rejected one utterance.  If you look in the training documentation, it says you should expect that up to a few percent of your utterances get rejected.

      What's happening (probably) is that your acoustic models don't match the training data for this utterance closely enough, and the Baum-Welch forward-backward computation is failing for that utterance.  This can happen due to noise in the data, an unusual pronunciation in the utterance, errors in the pronunciation dictionary, or just bad luck  (Gaussian mixture models have their problems).

      The system will ignore that one utterance for training, which is
      fine.  The error really is for diagnostic purposes only.  If you find that you are seeing this error for a large fraction of the training utterances, then you can start worrying.

      -Carl

       
      • Stephen Crowley

        Stephen Crowley - 2002-12-17

        Ahh, thanks for the help. How long should my utterences be? This one was about 30 seconds.

         
    • Carl Quillen

      Carl Quillen - 2002-12-17

      The documentation says utterances should be less than 100 seconds.   Apparently bw doesn't normalize every frame (I can't imagine why) and scores underflow after about 10000 frames.  In general long utterances will have an increased probability of rejection.

       
    • Carl Quillen

      Carl Quillen - 2002-12-17

      The documentation says utterances should be less than 100 seconds.   Apparently bw doesn't normalize every frame (I can't imagine why) and scores underflow after about 10000 frames.  In general long utterances will have an increased probability of rejection.

       

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.