Speech Recognition Engines

Flat
inaccurate recognition with Sphinx using voxforge-en
User: paulbunyan
Date: 7/5/2009 8:06 pm
Views: 4226
Rating: 6

I loaded voxforge-en 1.1 and 1.2 acoustic models into Sphinx, but I'm not getting very accurate results. It is better than the an4 tutorial am, but still not great. How can I improve accuracy?

Here's my cfgfile

-samprate 16000
-hmm /usr/src/sphinx/sphinx3/model/hmm/voxforge-en/model_parameters/voxforge_en_sphinx.cd_cont_3000
-dict /usr/src/sphinx/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp.sphinx.dic
-fdict /usr/src/sphinx/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp.sphinx.filler
-lm /usr/src/sphinx/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp_3gram.arpa.DMP

I also tried the voxforge_en_sphinx.cd_cont_untied, but Sphinx dies, saying I've exceeded the max senones.

--- (Edited on 7/5/2009 8:06 pm [GMT-0500] by paulbunyan) ---

Re: inaccurate recognition with Sphinx using voxforge-en
User: nsh
Date: 7/7/2009 9:34 am
Views: 1730
Rating: 6

Before you are going to improve something you need to measure it. Otherwise you won't notice the improvement. Probably there was a configuration mistake or something like. Also I suggest you to share your test data so we could check what are you doing.

--- (Edited on 7/7/2009 9:34 am [GMT-0500] by nsh) ---

PreviousNext