heard

What one piece comes back as, recovering an empty or prematurely stopped answer.

Parakeet answers some pieces of ordinary speech with no words at all, and whether it does turns on where the piece starts and how long it is together: the mel statistics are taken over the piece, so its length moves them, and past some edge the decoder predicts blank at every frame. Handing over a little less of the tail moves the piece off that edge. Nothing here tells speech from silence, so a piece that is genuinely silent pays for the whole list before answering nothing, which is why a piece shorter than shortest_worth_asking_again is not asked again at all.

A non-empty answer can also stop before speech resumes later in its audio. That tail is asked again with already recognised context and accepted only when the two answers share enough words to join without a duplicate.

An empty answer won by trimming is missing whatever was said in the tail that was cut off. Each piece overlaps the next, and that overlap is what covers it.