I've successfully build my custom acoustic model for a command and control application. I followed training acoustic model tutorial and tested with pocketsphinx like described.

For $CFG_HMM_TYPE  = '.semi.'; # Sphinx II configuration

I get Accuracy=98% in decoding test process with trained wav files and some new wav files not used in training process.

For $CFG_HMM_TYPE  = '.cont.'; # Sphinx III configuration

I get verry bad Accuracy=2% in decoding test process with the same test words and wav files.

I just rebuild acoustic model with new param's like was reported by SphinxTrain:

$CFG_HMM_TYPE  = '.cont.'

$CFG_STATESPERHMM = 3;

$CFG_FEATURE = "1s_c_d_dd";

What create such a big difference in generated models?

I need .cont model because I intend to use it with Sphinx4 in a Java application.

It is another way to use .semi model with Sphinx4 ? I have a very small dictionary (50 words) and I don't need continuous recognition, .semi model has acceptable results for me.

Best Regards!

--- (Edited on 1/25/2011 9:19 am [GMT-0600] by last_deceneu) ---