HTS Voice Training Checklist

Speaker-Independent Training

• Make sure you are using a template from the right version of HTS. Either copy from a similar voice, or use /proj/tts/hts-2.3/template_si_htsengine for HTS 2.3.
• Make sure the $MCDGV fix is present. This is a common training error for which we have embedded a fix into the training scripts. You can check that you have the fix by looking in yourvoicedir/scripts and seeing that a file called fix_mcdgv.py is there. If you used template_si_htsengine, then it should already be there. If not, please use a template that has it.
• Also under scripts, make sure that in Config.pm, $prjdir is set to the path to your voice directory.
• Acoustic data: under data/, put data from your corpus into directories cmp, lf0, and mgc. Typically we use symbolic links to avoid duplicating large amounts of data for each voice. If you are experimenting with different acoustic settings like f0 extraction ranges or other variations, then you'll have to re-make these with your own settings, rather than copying the data over.
• Text labels: under labels/, put data from your corpus into mono and full. Make sure that mono.mlf and full.mlf contain the right paths. If you are experimenting with some modified frontend, e.g. different pronunciations or syllabification or extra features, you will need to re-create these label files rather than using the default ones in the /proj/tts/data repository.
• Generation labels for synthesis: place these in labels/gen. Again, if you are using a nonstandard frontend, you will need to generate these labels yourself rather than using the default ones in the data repository.
• Model lists: these are lists/full_all.list, full.list, mono.list. full_all.list needs to contain labels for both training and synthesis, and also if you are using any non-standard frontend you will need to re-create these for the new model definitions resulting from your new labels. When in doubt, re-run this using make list in the data/Makefile (make sure to update any paths in there) after all your label files are in place.
• Training and generation script files. scp/train.scp is the list of files to use for training. Make sure they are all there, and that the paths to your voice directory are correct. If you are training a subset voice, this is the place to define the subset. scp/gen.scp lists the files to synthesize from for test synthesis. Make sure these contain the correct files and paths.
• Questions files under questions/. These are language-dependent so make sure you are using ones for the correct language. If you copied from template_si_htsengine then you are using the English ones. You can copy the ones for Turkish from any of the Turkish voices we have already trained. If you are using any of your own additional frontend features, then you will also need to add in questions about those in the questions file.
• Are you training using SAT? If so, then please also check the following:
  • Each data directory should have one subdirectory for each speaker.
  • gen labels should be prefixed with the ID of the target speaker and be in a subdirectory named after that speaker.
  • labels/*.mlf files should have speaker-specific lines.
  • scripts/Config.pm - check that $spkrPat is correct and also that $spkr is the ID of the adaptation speaker that you want to use.