Menu

Errors in training

Help
Anonymous
2008-03-01
2012-09-22
  • Anonymous

    Anonymous - 2008-03-01

    I am using Sphinx 3. I am getting diffenrent warnings and errors at different steps. I need to know what is the meaning each of these errors and why are they coming.

    In step 20.ci_hmm:
    ./hindi.1.2-1.bw.log:WARNING: "mod_inv.c", line 368: Model inventory n_density not set; setting to value in mixw file, 1.
    ./hindi.1.2-1.bw.log:WARNING: "mod_inv.c", line 256: n_top 4 > n_density 1. n_top <- 1
    ./hindi.1.2.norm.log:WARNING: NEGATIVE CONVERGENCE RATIO! CHECK YOUR DATA AND TRASNCRIPTS

    In 30.cd_hmm_untied:
    ERRORS:
    ./hindi.1-1.bw.log:ERROR: "backward.c", line 975: alpha(2.972315e-02) <> sum of alphas * betas (0.000000e+00) in frame 1233
    ./hindi.1-1.bw.log:ERROR: "baum_welch.c", line 331: hindi022 ignored
    ./hindi.2-1.bw.log:utt> 5 hindi006 999 0 372 13 ERROR: "backward.c", line 401: final state not reached
    ./t:./t:./hindi.3-1.bw.log:ERROR: "baum_welch.c", line 331: hindi054 ignored
    ./t:./t:./hindi.3-1.bw.log:utt> 54 hindi055 999 0 312 3 ERROR: "backward.c", line 401: final state not reached

    WARNINGS:
    ./hindi.1-1.bw.log:utt> 167 hindi171 899 0 236 20 WARNING: "gauden.c", line 1382: Scaling factor too small: -758.437908
    ./hindi.1.norm.log:WARNING: "gauden.c", line 1606: (mgau= 1114, feat= 0, density= 0) never observed

     
    • Nickolay V. Shmyrev

      I can only repeat:

      CHECK YOUR DATA AND TRASNCRIPTS

      Probably you are missing some silences in prompts and Baum-Welsh doesn't converge. Or you are just using broken dictionary.

       
      • ekelana

        ekelana - 2009-05-12

        Hi Nickolay

        I've got the "NEGATIVE CONVERGENCE RATIO" warning when doing SphinxTrain. Could you elaborate on "missing some silences in prompts" ?
        About using broken dictionary, I have checked my dictionary and i think it's just fine. The verify_all.pl scripts didn't give any warning or error.

        regards
        Eka

         
        • Nickolay V. Shmyrev

          I don't quite get what is your trouble in understanding this. The training process diverge because the phonetic transcription extracted from your transcriptoin file and your dictionary is not well suitable for the actual recording you made.

          For example there is a silence in a recorded audio and it's not present in transcription file. Or all your speakers say HELLO as H EH L OW while your dictionary has the entry

          HELLO G UH D B AY

          verify_all doesn't check that obviously.

           

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.