Hello,

After running upto module 5 correctly, I'm having trouble running module 6  "mllt_train" of the sphinx3 trainer. When i run the script "06.mllt_train/slave_mllt.pl" i get the following error in my "06.mllt_train/dkk3.initvar_cihmm.log" file

INFO: init_gau.c(157): Computing 1x1x1 variance estimates
WARN: "s3io.c", line 256: Unable to open /opt/dkk3_mllt_final/model_parameters/dkk3.ci_mllt_flatinitial/globalmean for reading; No such file or directory
FATAL: "init_gau.c", line 166: Unable to open /opt/dkk3_mllt_final/model_parameters/dkk3.ci_mllt_flatinitial/globalmean for reading
Fri Jun 12 15:50:46 2009

After some debugging we traced the problem to the binary file init_gau which gives the error "segmentation fault" when executed with the following paramters

-ctlfn /opt/dkk3_mllt_final/etc/dkk3_train.fileids -part 1 -npart 1 -cepdir /opt/dkk3_mllt_final/feat -cepext mfc -accumdir /opt/dkk3_mllt_final/bwaccumdir/dkk3_buff_1 -agc none -cmn current -varnorm no -feat 1s_c -ceplen 13 -ldafn /opt/dkk3_mllt_final/model_parameters/dkk3.lda -ldadim 2

All the files that we pass here do exist. I would really appreciate any information which can help us resolve the above issue as it's driving me crazy!

--- (Edited on 12-June-2009 4:43 pm [GMT+0500] by ersalan) ---

> I would really appreciate any help which can help us resolve the above issue as it's driving me crazy!

It's quite crazy to set ldadim to 2, are you sure you need it?

And feat 1s_c also bad, do you understand what did you change in config?

--- (Edited on 6/12/2009 6:40 am [GMT-0500] by nsh) ---

Thanks for the quick reply! Really appreciate it. Urgh how could i be such a dunce! You were right by pointing out those two configuration variables. Changed those and things have moved forward now. Am still facing a few issues but will try and resolve them myself first with the help of other links on this forum.

Many Thanks! : )

--- (Edited on 12-June-2009 5:01 pm [GMT+0500] by ersalan) ---

I've encountered another problem that's proving to be a real pain. When I run the RunAll.pl script, everything runs peachy however when i run the RunAll_CDMLLT.pl script, it breaks on module 47. The weird thing is that when I run module 47 (mllt train) individually after running RunAll.pl, it works fine but it only breaks after i run module 46 (lda train). I don't see any problems with module 46 itself as that does not generate any serious error. After some debugging we came to know that the training breaks on the following command

bw -moddeffn /opt/dkk3_mllt_final/model_architecture/dkk3.1000.mdef -ts2cbfn .cont. -mixwfn /opt/dkk3_mllt_final/model_parameters/dkk3.lda_cont_flatinitial/mixture_weights -mwfloor 1e-05 -tpfloor 1e-05 -tmatfn /opt/dkk3_mllt_final/model_parameters/dkk3.lda_cont_flatinitial/transition_matrices -meanfn /opt/dkk3_mllt_final/model_parameters/dkk3.lda_cont_flatinitial/means -varfn /opt/dkk3_mllt_final/model_parameters/dkk3.lda_cont_flatinitial/variances -ltsoov no -dictfn /opt/dkk3_mllt_final/etc/dkk3.dic -fdictfn /opt/dkk3_mllt_final/etc/dkk3.filler -ctlfn /opt/dkk3_mllt_final/falignout/dkk3.alignedfiles -part 1 -npart 1 -cepdir /opt/dkk3_mllt_final/feat -cepext mfc -lsnfn /opt/dkk3_mllt_final/falignout/dkk3.alignedtranscripts -accumdir /opt/dkk3_mllt_final/bwaccumdir/dkk3_buff_1 -varfloor 0.0001 -topn 4 -abeam 1e-90 -bbeam 1e-10 -agc none -cmn current -varnorm no -meanreest 'yes' -varreest 'yes' '-2passvar' no -fullvar no -diagfull no -feat 1s_c_d_dd -ceplen 13  -ldafn /opt/dkk3_mllt_final/model_parameters/dkk3.lda -ldadim 29 -timing 'no'

The error it generates is

bw: feat.c:268: feat_read_lda: Assertion `lda.lda_cols == feat_conf[fid].blksize()' failed.
Aborted

Can anybody tell me what I'm doing wrong here.

--- (Edited on 15-June-2009 3:44 pm [GMT+0500] by ersalan) ---

dont run this script. it was actually removed recently in a released version.

--- (Edited on 6/15/2009 5:58 am [GMT-0500] by nsh) ---

Cool! Thanks again for the quick reply. If possible can you direct me to the latest build where this script is not used anymore.

--- (Edited on 15-June-2009 5:02 pm [GMT+0500] by ersalan) ---

http://cmusphinx.sourceforge.net/html/download.php#SphinxTrain

--- (Edited on 6/15/2009 4:57 pm [GMT-0500] by nsh) ---

Thanks! My last query regarding training is as follows.

In the sphinx_train.cfg file on line 133 we have a variable "CFG_FORCE_ALIGN_MDEF" defined with comments that hint that this variable can be modified to use other model deffinitions. If this is the case then i was wondering if anybody can let me know about what other options i have for this variable and what other variables i may need to update if this variable is modified. Like for example will i be gaining anything by using the untied model definition instead of the default falign_ci model definition?

--- (Edited on 16-June-2009 1:59 pm [GMT+0500] by ersalan) ---

> If this is the case then i was wondering if anybody can let me know about what other options i have for this variable and what other variables i may need to update if this variable is modified.

Obviously you can use any other model with the same phoneset there :) I suggest you to ask more exact questions.

>Like for example will i be gaining anything by using the untied  model definition instead of the default falign_ci model definition?

No. Untied model is too bounded to training data because it has all triphones in the database, so it's very bad for alignment. There is little sense in using cd model but usually it's just slow without much gain. The real use could be in using another completely different model with very good quality to align a bad data and train on it, like use hub4 to align wsj, but it's not your case I suppose.

--- (Edited on 6/16/2009 1:07 pm [GMT-0500] by nsh) ---