Hello, I would like to ask how to restrict Julius to try to match every noise to my dictionary (my.voca)? Each time I say something (even in czech language), Julius simply outputs CALL HELLO or TEST or something. When I actually say the phrase CALL HELLO or CALL OPERA, Julius detects this correctly, with scoring 1.0 1.0 1.0. But the wrong phrase is often 1.0 as well or 1.0 0.8 1.0.

My goal is to control the Linux desktop and I need ideally one phrase to switch on recognition and then say command. This I'am able to do on myself but I need Julius to react on specific word, not everything.

I've tried to adjuct mic level and various settings, but did not find a glue. I'am using PS3 Eye camera, have 35 lines of 48Khz prompts and adapted the model using the script HTK_Compile_Model.sh.

Any help would be much appreciated.

My .gram

S : NS_B SENT NS_E

 SENT: WHO_V SAY

 SENT: WHO_V DO

 SENT: TEST_V

and .voca

%

% NS_B

 <s>            sil

% NS_E

 </s>           sil

% WHO_V

 CALL        k ao l

% SAY

 HELLO          hh ax l ow

 PLAY           p l ey

 SIGN           s ay n

% TEST_V

 TEST           t eh s t

% DO

 OPERA          aa p r ax

 WEATHER        w eh dh er

 NEWS           n uw z

julius.my_grammar.conf

-v my.dict

-h acoustic_model_files/hmmdefs

-hlist acoustic_model_files/tiedlist

-smpFreq 48000

-dfa my.dfa

-quiet

-cutsilence

-realtime

-progout

--- (Edited on 9/2/2011 5:03 am [GMT-0500] by Visitor) ---