Menu

Scaling factor small error log file RunAll.pl

Help
2010-10-26
2012-09-22
  • Gaurav Arora

    Gaurav Arora - 2010-10-26

    we were trying to build a acoustic model on collection of data (speech wav
    file) using http://cmusphinx.sourceforge.net/wiki/tutorialam .

    1. while creating acoustic model i created dictionary ,filler dictionary ,fields file and transcript all file required for the acoustic model training
    2. setting acoustic model using scripts(setup_SphinxTrain.pl) in sphinx train.
    3. then creating feature for the wav file using make_feat.pl script in SphinxTrain.
    4. running ./scripts_pl/RunAll.pl to train model

    Html log file created by sphinx Train show warning and errors relating to the
    training acoustic model:

    Phase 3: Forward-Backward

    Baum welch starting for 1 Gaussian(s), iteration: 1 (1 of 1)

    bw Log File

    This step had 28 ERROR messages and 12 WARNING messages. Please check the log
    file for details.
    **
    ERROR MESSAGE

    utt> 13 FM014014 104 0 20 12 WARNING: "gauden.c", line 1370: Scaling factor
    too small: -873.324838
    ERROR: "backward.c", line 1171: alpha(3.315119e-05) <> beta(3.588999e-81)
    1 3 1.423793e-12ERROR: "baum_welch.c", line 331: train/FM014014 ignored

    **same error is repeating with 14 annotation out of 4494.
    and error repeats in each iteration.

    whereas after completion of training i am able to run pocketsphinx_continous
    ,batch commands with new hmm model created by this and also give results.

    Is this problem with respect to speech files as only some of speech files are
    ignored.Is there any way to remove these errors .
    Error is related to alpha and beta values calculated using speech file.Does
    ignoring this error hamper performance of acoustic model.

    I will be obliged if some one can point me resources which can let me know how
    small values of alpha and beta leads to ignoring features of these speech
    file.

     
  • Nickolay V. Shmyrev

    Is this problem with respect to speech files as only some of speech files
    are ignored

    Yes

    Is there any way to remove these errors.

    Probably files have noise, zero energy regions or some other issues. Remove
    those files from training database or fix their transcription. Or record them
    once again.

    Does ignoring this error hamper performance of acoustic model.

    No

    I will be obliged if some one can point me resources which can let me know
    how small values of alpha and beta leads to ignoring features of these speech
    file

    Rabiner HMM Tutorial
    http://www.cs.ubc.ca/~murphyk/Bayes/rabiner.pdf

     
  • Gaurav Arora

    Gaurav Arora - 2010-10-26

    thanks...

     

Log in to post a comment.