MLLT training Problem

[CAO.Z.H]

Following http://www.speech.cs.cmu.edu/cmusphinx/moinmoin/LDAMLLT
I've trained a MLLT model.

But i can't find 'feature_transform' lives inside the acoustic model directory.
What is the problem?

Thanks

[CAO.Z.H]

I still can't find the 'feature_transform' file. After MTTL training, i've got following floders under model_parameters

test.cd_cont_1000 test.cd_cont_1000_2 test.cd_cont_1000_8 test.cd_cont_untied test.ci_cont_flatinitial test.lda_cont_1000
test.cd_cont_1000_1 test.cd_cont_1000_4 test.cd_cont_initial test.ci_cont test.lda test.lda_cont_flatinitial

Under each folder, there are files:
./test.cd_cont_1000:
feat.params mdef means mixture_weights noisedict transition_matrices variances

./test.cd_cont_1000_1:
means mixture_weights transition_matrices variances

./test.cd_cont_1000_2:
means mixture_weights transition_matrices variances

./test.cd_cont_1000_4:
means mixture_weights transition_matrices variances

./test.cd_cont_1000_8:
means mixture_weights transition_matrices variances

./test.cd_cont_initial:
means mixture_weights transition_matrices variances

./test.cd_cont_untied:
feat.params mdef means mixture_weights noisedict transition_matrices variances

./test.ci_cont:
feat.params mdef means mixture_weights noisedict transition_matrices variances

./test.ci_cont_flatinitial:
globalmean globalvar means mixture_weights transition_matrices variances

./test.lda_cont_1000:
feat.params mdef means mixture_weights noisedict transition_matrices variances

./test.lda_cont_flatinitial:
globalmean globalvar means mixture_weights transition_matrices variances

I still can't find 'feature_transform' file,
what is the problem ?

[CAO.Z.H]

Anybody who can tell me?
thank a lot

[Nickolay V. Shmyrev]

Missing python probably? The reason must be hidden somewhere in the log. The corresponding part of the script which creates the file:

my $mlltfile = catfile($ST::CFG_MODEL_DIR, "${ST::CFG_EXPTNAME}.mllt");
if (-r $mlltfile) {
    copy($mlltfile, catfile($hmmdir, 'feature_transform'));
}

[David Huggins-Daines]

Make sure you are looking in the right directory. It is only the "final" acoustic model directory that gets the feature_transform file. This is the one that doesn't have the number of Gaussians appended to it.

So for example if you are training wsj with 5000 senones and 16 Gaussians, it is NOT wsj.mllt_cd_cont_5000_16, but rather wsj.mllt_cd_cont_5000.

(also note that the final model for MLLT is wsj.mllt_cd_cont_5000, not wsj.cd_cont_5000)

[CAO.Z.H]

I failed in getting final AM.
Here is the error message:
===================================================================================
Phase 3: Forward-Backward
Baum welch starting for MLLT, iteration: 1 (1 of 1)
0%
Only 0 parts of 1 of Baum Welch were successfully completed
Parts 1 failed to run!
Something failed: (/home/zhenhai/training/test/scripts_pl/47.mllt_train/slave_mllt.pl)

here is the error message in lof file logdir/47.mllt_train/test.1-1.bw.log

bw: feat.c:268: feat_read_lda: Assertion `lda_cols == feat_conf[fid].blksize()' failed.
Wed Mar 26 17:29:07 2008

what's the problem?

thanks

[CAO.Z.H]

Anybody who can tell me?
thank a lot

[Nickolay V. Shmyrev]

Well, it would be helful to know the values of variables in assertion. You can get them inserting additional prinf there. Then one can look where this variables are assigned, probably configuration file is not complete or not consistent.

Is it possible to reproduce the problem for me?

[CAO.Z.H]

Thanks for your advice.
I've fix this problem. The reason is i add a new filler (besides SIL) in file test.phones and test.filler, while it not appeared in transcription.
But when i use ordinary training, it is OK.