On Jul 31, 2009, at 8:18 AM, Helmut Kuper wrote:
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
Hello,
I spent a few days working on my problem. I played around with voxforge's data, read some of the train docs of CMU, and debugged the sphinxbase, pocketsphinx and mod_pocketsphinx.
Results: - -I found a way to use the voxforge data as training data for creation of a german language corpus.
Care to document the process?
- -I enabled the logging of pocketsphinx to stderr (Dirty, but easy way to see what went wrong when FS loads grammar, mdef, etc and simply stopped. Very helpful!
I can't recall if there is a logger callback we can register for this... Last I checked you couldn't this is something we should make a config option for up in the mod if possible.
- -I had to add a "dictcase" parameter to pocketsphinx.conf.xml resp mod_pocketsphinx.c to allow case sensitive dictionaries (like the german dictionary from voxforge).
What do you mean? Can you put this on jira please.
FS starts up with german language model and detects the words as expected. But it's not so reliable as I want to ... I guess this is caused by the very small amount of training audio data. I used 4000 of 19000 audio files provided by voxforge due to the reason that voyforge's training fileid-list contains only 4000 files ... I have to create new fileid-list and transcription-lists containing all audio I have downloaded from voxforge.
Quite complex the whole thing ...
regards helmut