Hi degra,

>voxforge AM with gigaword LM,

we've got a long way to go before we can do dictation... (much more speech required).

>I wonder if it is due to my accents and how to improve that?

You might want to try adapting the VoxForge acoustic model to your voice.

Please  post the steps you took to get the VoxForge AM and the Gigawod LM together,

thanks,

Ken

--- (Edited on 9/21/2009 3:10 pm [GMT-0400] by kmaclean) ---

Hi Ken,

I'm sorry i never mentioned i'm using sphinx3 in my first post.

I did follow this link http://www.speech.cs.cmu.edu/cmusphinx/moinmoin/AcousticModelAdaptation to exactly adapt my own voice.

These are the command i use to adapt my voice.

./bw     -hmmdir /root/sphinx/voxforge-en/model_parameters/voxforge_en_sphinx.cd_cont_3000     -ts2cbfn .cont. -feat 1s_c_d_dd -cmn current -agc none -dictfn me72.dic     -fdictfn me72.filler     -ctlfn me72.listoffiles     -lsnfn me72.transcription -accumdir .

./mllr_solve     -meanfn /root/sphinx/voxforge-en/model_parameters/voxforge_en_sphinx.cd_cont_3000/means     -varfn /root/sphinx/voxforge-en/model_parameters/voxforge_en_sphinx.cd_cont_3000/variances     -outmllrfn mllr_matrix -accumdir .

and these are the parameters used to run the sphinx3_livepretend

-samprate 16000
-hmm /root/sphinx/voxforge-en/model_parameters/voxforge_en_sphinx.cd_cont_3000
-dict /root/sphinx/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp.sphinx.dic
-fdict /root/sphinx/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp.sphinx.filler
-lm /root/sphinx/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp_3gram.arpa.DMP
-mllr /root/sphinx/adapt/me72/mllr_matrix

In fact i would like to give julius with LM a try, unfortunately even after i've read the juliusbook, i still don't know where to start off.

Regards

--- (Edited on 9/22/2009 9:28 am [GMT-0500] by degra) ---

--- (Edited on 9/22/2009 9:29 am [GMT-0500] by degra) ---

--- (Edited on 9/22/2009 9:31 am [GMT-0500] by degra) ---

voxforge 0.1.2 is a model with feature transform. You need to pass -ldafn voxforge-en/model_parameters/voxforge_en_sphinx.cd_cont_3000/feature_transform

argument to bw. Also you most probably ignored the errors during the adaptation that appeared due to that. I suggest you to show your adaptation data if you need more help.

--- (Edited on 9/22/2009 18:47 [GMT+0400] by nsh) ---