False detection for Pocketsphinx on Android - English language

[paul]

Hello everyone,

I am working on a project where I have to integrate the speech functionalities of Pocketsphinx into an Android application.

I have to detect 3 action words of which I've made a small grammar and a small dictionnary generated with LMTOOL (above a dozen words).

My grammar :

#JSGF V1.0;

grammar menu;

public <action> = VALIDATE | BACK | RETURN ;

I'm using the following assets :
* dictionnary : custom
* Language Model : custom
* Acoustic Model : latest en-us-ptm

Actually I'm having a lot of false detection problem. For example a small noise is recognize as "back", or "validate" is recognize as "return". I don't know if the problem comes from the Android device microphone or a bad configuration (probably more this one), but I've seen there are a lot of pocketsphinx config parameter and I don't know which one are useful or not.

I've only configure theses :

             .setKeywordThreshold(1e-45f)

             .setBoolean("-allphone_ci", true)

             .setFloat("-lw", 6.5)

I've made a sample raw with a noise which is detected and one with the "validate" command.

So, could you guys please help and let me know what could be the problem and any suggestions for solving this problem?

Thank you very much!

Paul

[Nickolay V. Shmyrev]

To recognize words in continuous speech you need to use keyword spotting mode, not grammar mode as covered in tutorial

http://cmusphinx.sourceforge.net/wiki/tutoriallm

[paul]

Hello Nickolay,

Thank you very much for your fast response.

I've already tried but with addKeywordSearch the recognition is near 0, like the decoder is not enough sensitive (inverse problem to the previous post).

Best,

Paul

[Nickolay V. Shmyrev]

Dear Paul

Please read the tutorial linked above about tuning the detection threshold.

[paul]

Hello again Nickolay,

I've followed your advices and completed the tutorial (see the attachtment) but is there a script like "word_align.pl" for the Tuning Tutorial to transcript the result in a more understable form ?

Best,

Paul

[Nickolay V. Shmyrev]

No, there is no script yet

[paul]

Hello Nickolay,

Thank you very much for your previous response, i'll find a way around with counting results...I 've got another question : in this case is the threshold the only parameter we can play with to improve the performances or there are others ?

Best,

Paul

[Nickolay V. Shmyrev]

You can use longer spotting phrase.