Menu

Normalization error when training

Help
roof
2005-12-08
2012-09-22
  • roof

    roof - 2005-12-08

    Hello,

    I am training using SphinxTrain on linux, a small sub-sample (200 audio files) of my data in order to test that it works so that I can apply the same training over the entire dataset (4000+ audio files).

    When training I recieve a stream of errors during the Normalization for Iteration 1 step when training Context Dependant Models and Context Independant Models.
    ERROR: ERROR: "gauden.c", line 1418: var (mgau= 1082, feat= 0, density=7, component=9) < 0

    ERROR: ERROR: "gauden.c", line 1418: var (mgau= 1082, feat= 0, density=7, component=16) < 0

    ERROR: ERROR: "gauden.c", line 1418: var (mgau= 1082, feat= 0, density=7, component=18) < 0

    ERROR: ERROR: "gauden.c", line 1418: var (mgau= 1082, feat= 0, density=7, component=19) < 0

    There are literally hundreds of these errors, and I have no idea what they mean..
    This is after the baum welch step for interation 1 has completed (usually without error).
    They only seem to occur during the FIRST iteration, and on subsequent iterations the logfiles show the following types of errors:

    WARNING: "gauden.c", line 1376: (mgau= 2886, feat= 0, density= 0) never observed
    WARNING: "gauden.c", line 1376: (mgau= 2887, feat= 0, density= 0) never observed

    Has anyone observed similar errors and know where to start fixing them?

     
    • roof

      roof - 2005-12-08

      Ah, not to worry, after going through the forums once again I found a similar thread replied to already that said that this error is caused by having a very small training set

      Since I have a large set available to me, I intend to try it over the whole lot, but can anyone confirm that these issues are indeed cleared up by including over 10 hours of audio data?

      I am a little worried about, taking the days/weeks to train the whole lot only to find the error still persists!

      Cheers!

       

Log in to post a comment.