I’m envious of your partner; Bozen is a lovely place. OK, so does that mean that we can say that eventually recognition is performed when a very long sound stops, but it is unpleasantly slow because it is finally scanning all of what has become an extremely long utterance? In that case, maybe you can get your solution from this thread about a similar issue, in which we found a workaround: https://www.politepix.com/forums/topic/cancel-current-processing-on-pause-detected/
I might end up adding this feature to the main distribution since it seems like in keyword spotting mode, few developers want such long utterances to be completely analyzed.