Menu

help with Errors returned by sphinx

Help
abs
2006-09-15
2012-09-22
  • abs

    abs - 2006-09-15

    hi,

    we were training the sphinx recognizer using features (not mfcc) when we got this error:

    <b>
    bw: gauden.c:1425: gauden_scale_densities_bwd: Assertion `finite(den[c][j][k])' failed.
    </b>

    and this warning

    <b>
    WARNING: "gauden.c", line 1418: Scaling factor too small: -913.889597
    ERROR: "backward.c", line 975: alpha(1.478608e-01) <> sum of alphas * betas (0.000000e+00) in frame 206
    ERROR: "baum_welch.c"
    </b>

    Please shed some light on this. Thanks in advance.

     
    • David Huggins-Daines

      Hi,

      These errors usually mean either that something is pathological about your training data, or that there is a mismatch between the model used to run forward-backward and the data on which you are running it.

      The assertion has been triggered (as the message indicates) because of an infinity in the Gaussian density calculation. This could occur if the variance for some mixture is zero.

      Could you say a bit more about the features you're using?

       
    • abs

      abs - 2006-09-22

      thanks !

      we are testing some new features developed in our lab.

      rgds

       

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.