[hts-users:01268] HERest Warning Message

["Tillman, Raymond T CTR USAF AFMC AFRL/RHA" <raymond.tillman@xxxxxxxxxxxxxxxx>]

I am a newby in HTS. I am having a problem with a voice that I have created using the HTS-2.0.1 for HTK 3.4 patched version of HTK. 

I used the automated build_hts script and built a voice from some 511 utterances based on the arctic database. Both the festvox/*.scm and hts/* parameter files were generated. There were warnings generated while building the voice. The voice does not work in Festival. I get an malloc <negative number> exception and festival dies. Recognizing that there may be variances between the voices built for Festival and the newer v2.0.1 HTS, I also ran with the hts_engine that is included in the HTS-2.0.1 patch to HTK. That engine also fails and reports:

       HTS_HTS_ModelSetLoad: voiced/unvoiced weights must be within .99 to 1.01.

There were warnings when I built the voice. The HTS training script marches through just fine until it completes iteration 5 of embedded Re-estimation. After the application reports processing data for the last label file for my audio dataset, the application begins sending many thousands of warning messages regarding the error 7023 instead of what it does normally. The error message:

      WARNING [-7023] SetVFloor: the varFloor# is not a variance floor macro in /usr/local/HTS-2.0.1/bin/HERest
      
The count in '#' cycles between 1 and 4.

Then, it reports (as it is supposed to do):

      Total 9769 floored variance elements in 2170 different mixes
      Saving hmm's to dir /home/festival/festival/lib/voices/us/afrl_us_rfs_arctic_hts/hts_build/train/hmms/db/v/cmp
      Total 9769 floored variance elements in 2170 different mixes
      Saving duration models to MMF 
      /home/festival/festival/lib/voices/us/afrl_us_rfs_arctic_hts/hts_build/train/hmms/db/v/dur/fullcontext_dur.mmf
      Reestimation complete - average log prob per frame = 1.039788e+02
           - total frames seen          = 3.838600e+05
      ==================================================
      Realtime 0: 1.27
      ==================================================

      ==================================================
      -rw-r--r-- 1 ray None 14413436 Mar 25 14:05     
      /home/festival/festival/lib/voices/us/afrl_us_rfs_arctic_hts/hts_build/train/hmms/db/v/cmp/clustered_cmp_qst001.mmf
      ==================================================


       ==================================================
       Tue Mar 25 14:05:51 MST 2008
       ==================================================

       HTS training: Duration Model Clustering

After this, the model continues to build.

I have made no changes to the Training.pl script.

When I tried to run the mkdata_db_v.pl script, I got the error: "the log periodgram was 0 use -e option."

When I added the option -e <value>, the voice built but I could not use the voice. I do not know what the problem is. I have tried a very small value (.001) and a larger value (1.0) with no effect on the above. I have traced the HERest code but only learned WHERE the problem happens, not why.

I don't know what the problem is. It seems to me that the voice did not build correctly, because of the warnings. 

Please give me a little help with this problem. 

Also, are voices built with the HTS 2.0.1 patch compatible with Festival? If not, is there a way to use Festival for the front end and use the hts_engine as a back end?

Dr. Raymond T. Tillman
L-3 Communications, Inc., Link Simulation & Training
Raymond.T.Tillman@xxxxxxxxxx
Research Scientist, AFRL/RHA
Raymond.Tillman@xxxxxxxxxxxxxxxx
US Air Force Research Labs
Arizona State University, Polytechnical Campus
6030 South Kent, Mesa, AZ 85212
(480) 988-9773 x238
 
Caution: This message may contain competitive, sensitive or other non-public information not intended for disclosure outside official government channels. Do not disseminate this message without the approval of the undersigned's office. If you received this message in error, please notify the sender by reply e-mail and delete all copies of this message.