Following Sphinx Tutorial

[Rossi Luo]

When I followed the "How to perform a preliminary training run" step and do
the "perl scripts_pl/RunAll.pl
" action, it failed at 20.ci_hmm/slave_convg.pl. The description in an4.html
is

MODULE: 20 Training Context Independent models (2010-02-11 13:45)

Phase 1: Cleaning up directories:
accumulator... logs... qmanager... models... completed

Phase 2: Flat initialize

mk_mdef_gen Log File
completed

mk_flat Log File
completed

Phase 3: Forward-Backward

Baum welch starting for 1 Gaussian(s), iteration: 1 (1 of 2)

bw Log File

Baum welch starting for 1 Gaussian(s), iteration: 1 (2 of 2)

bw Log File

This step had 950 ERROR messages and 1852 WARNING messages. Please check the
log file for details.
completed

This step had 946 ERROR messages and 1841 WARNING messages. Please check the
log file for details.
completed

Training failed in iteration 1

I am using cygWin in Windows 7.

I have searched the forum, but nothing useful is found. Many thanks for any
help.

[Nickolay V. Shmyrev]

This step had 950 ERROR messages and 1852 WARNING messages. Please check the
log file for details.

It tells you to check log for details. Did you try to follow this wise
suggestion? Logs are in logdir subfolder in an4 training folder.

[Rossi Luo]

As recorded in the logs all the errors are from this:

ERROR: "backward.c", line 1020: alpha(2.500000e-01) <> sum of alphas * betas
(0.000000e+00) in frame 97
ERROR: "baum_welch.c", line 331: fash/an251-fash-b ignored

The responding code is

/* check an invariant.  Theoretically,
     * sum_alpha_beta - alpha[n_obs-1][n_state-1] must be zero, but
     * we're dealing with finite machine word length, pruning, etc. */

    if (fabs(sum_alpha_beta - active_alpha[n_obs-1][q_f])
        > (S2_ALPHA_BETA_EPSILON * active_alpha[n_obs-1][q_f])) {

        E_ERROR("alpha(%e) <> sum of alphas * betas (%e) in frame %d\n",
            active_alpha[n_obs-1][q_f], sum_alpha_beta, t);

        retval = S3_ERROR;

        goto free;
    }




    ckd_free((void *)active_alpha);
    ckd_free((void *)active_astate);

    E_ERROR("%s ignored\n", corpus_utt_brief_name());

    return S3_ERROR;

I don't understand the code, thanks very much for you help。

Since I followed the tutorial exactly, I am not supposed to meet such a
problem, am I?

Many thanks

[Nickolay V. Shmyrev]

Its most likely that your files were currupted during extraction. I remember
cygwin has special mode when on extraction it automatically adds windows-style
CR+LF to the files. Check md5sum of an251-fash-b.sph, it should be
b590c787769e6d13cb8ccc693c57b303 an251-fash-b.sph.

[Nickolay V. Shmyrev]

And, cygwin is a pain for various reasons as it always was. Install linux in a
virtual machine and enjoy real speech recognition development environment.

[Rossi Luo]

Thanks nshmyrev, I dropped cygwin and installed vmware and ubuntu, then
followed the tutorial's instructions again. Yes, all the perl scripts seems
excuted properly. But I havn't got the "./model_parameters/an4.1000.s2models"
which is seems available immediate after the training finished according to
the tutorial.

This is the log file:
http://docs.google.com/View?id=dd9b6sxp_123fv4nqjc9

Many thanks.

[Nickolay V. Shmyrev]

Right, since you trained continouos model result will be in an4.cd_cont_1000.
See earlier in tutorial

$CFG_HMM_TYPE = this variable could have the values .semi. or .cont.. Notice
the dots "." surrounding the string. Use .semi. if you are training semi-
continuous HMMs, mostly for Pocketsphinx, or .cont. if you are training
continuous HMMs (required for SPHINX-4, and the most common choice for
SPHINX-3)