Keywords spotting: hypotesis is null until it is exactly one of the keyword you are looking for

I'm using onPartialResult method to look if the hypotesis is one of the keyword I'm interested in and it works well.

Here's my code:

@Override
public void onPartialResult(Hypothesis hypothesis) {
    Log.d(TAG, "onPartialResult");

    if (hypothesis == null) {
        return;
    }

    String text = hypothesis.getHypstr();

    String wordWithProb = "";
    String mostProbableWord = "";
    int probability = -10000;

    if (text.contains("|")) {
        for (Segment seg : recognizer.getDecoder().seg()) {
            wordWithProb += "|" + seg.getWord() + " " + seg.getProb() + "|";
            if (seg.getProb() > probability)
                mostProbableWord = seg.getWord().trim();
        }
    }
    else
        mostProbableWord = text.trim();

    Log.i(TAG, "onPartialResults: " + mostProbableWord);

    String recognizedCommand = "Please repeat";
    if (mostProbableWord.equals("one")) {
       //do something...
    } else if (mostProbableWord.equals("two")) {
       //do something...
    } else if (mostProbableWord.equals("three")) {
       //do something...
    } 

    //text to speech
    speak(recognizedCommand);

    startListening(KWS_SEARCH);
}

Now I would like to handle the case where a user say something and it is not recognized as a keyword; in this case the hypotesis in the onPartialResult method is always null: is this expected? I was expecting a not null hypotesis here...
Considering that onPartialResult method is automatically called by pocketsphinx consinuously (also when there isn't any sound in the air) I cannot use the null hypothesis as my driving condition.
Moreover there is a text to speech after every recognition and so the recognition listener restart has to be handled carefully: recognizer must not be listening while text to speech is ongoing...
I tried some solution with onEndOfSpeech but none was good until now... Any idea?

in this case the hypotesis in the onPartialResult method is always null: is this expected?

Yes

Moreover there is a text to speech after every recognition and so the recognition listener restart has to be handled carefully: recognizer must not be listening while text to speech is ongoing

Correct