JSGF Grammar Partly Ignored

[Cubeler]

Hello,

I'm just getting started with sphinx and adopted the transcriber demo to use my own grammar:

#JSGF V1.0;
grammar test; 
< item > = (Some Else) | Keyboard;
public <command> = add (< item >)+ end;

This works as expected and the hypothesis i get is: add keyboard end.

Now the strange part.
When i change the item line to:

<item> = Vantage | Keyboard;

The hypothesis is: add keyboard. So the end is missing.

This shouldnt happen since the "end" is mandatory, right?
Thanks for your help

[Nickolay V. Shmyrev]

This shouldnt happen since the "end" is mandatory, right?

No

JSGF describes a sequence of words, not the full recognition result.

[Cubeler]

So how do i force to only consider the whole rule as a result?

[Nickolay V. Shmyrev]

So how do i force to only consider the whole rule as a result?

It depends on sphinx4 version you are using

[Cubeler]

sphinx4-5prealpha

[Nickolay V. Shmyrev]

There is no support for that in the api yet, you have to edit sources. Change the following in edu.cmu.sphinx.api.SpeechResult:

public String getHypothesis() {
    return result.getBestResultNoFiller();
}

to

public String getHypothesis() {
    return result.getBestFinalResultNoFiller();
}

That will force recognizer to return final result.

[Cubeler]

After this change i don't get any hypothesis in the second case.

[Nickolay V. Shmyrev]

Right because recognizer was unable to find final hypothesis in recognition result due to inaccurate recognition. To get help on accuracy you need to provide the test data you are trying to decode.

[Cubeler]

Is it possible to get the most likely result although the accuracy is very low?

[Nickolay V. Shmyrev]

You already got most likely result and this result is empty. Somehow probability of the empty result was significantly higher than the probability of any final result. To get suggestions on that you need to provide an audio. There are also more accurate models in our downloads like en-us generic.

[Cubeler]

Okey that makes sense.
The audio file is pretty unaccurate to test the behaviour since I can only return a valid sentence of the grammar.
The accurate files are recognized just fine.

Thanks for your help