keyword spotting with pocketsphinx.

[creative64]

Hi,

Inorder to get a simplistic keyword-spotting solution with pocketsphinx, I
tried with a simple grammar file. It seems
to give OK results for applications where robustness needn't be very-very
high. Would like to get some feedback on
performance of this approach.

I've tried it with model hub4wsj_sc_8k with pocketsphinx ver 0.6.1. Please
note that bestpath and nbest option need to be
turned off otherwise it takes a very long time to decode.

My keyword is "HELLO SIMSIIM" and grammar file is as follows:

#JSGF V1.0;  

grammar kewword_test;  

public <command> = <option_1> | <option2> | <option3> | <option4> | <option5> ;  

<option_1> = <keyword>;  

<option_2> = <garbage_loop> <keyword>;  

<option_3> = <keyword> <garbage_loop>;  

<option_4> = <garbage_loop> <keyword> <garbage_loop>;  

<option_5> = <garbage_loop>;  

<keyword> = (HELLO SIMSIM);  

<garbage_loop> = (G1 | G2 | G3 | G4 | G5 | G6 | G7 | G8 | G9 | G10 | G11 | G12
| G13 | G14 | G15 | G16 | G17 | G18 | G19 | G20 | G21 | G22 | G23 | G24 | G25
| G26 | G27 | G28 | G29 | G30 | G31 | G32 | G33 | G34 | G35 | G36 | G37 | G38
| G39)+ ;

[creative64]

and.....G1 to G39 are mapped to 39 base phonemes of the model in dictionary.

[Nickolay V. Shmyrev]

Would like to get some feedback on performance of this approach.

It's ok, more or less what you need. You need to care about garbage path
pruning here, though it's easier to be done in kws-specific code. We have our
own KWS implementation in Nexiwave based on pocketsphinx, but it's different
ps_search class, not fsg search.

[Biswajit Das]

I had tested with garbage model like, added all phoneme into dictionary and also in the JSGF grammar. Only phoneme being recognized besides words. fsg_search taking lot of time and hanging the system. Should I, reduce number of phone in garbage?

[Anonymous]

You probably should reduce the length of the processed
audio, accounting for shorter silences and process shorter amount of speech.