Vocal Burst Segments — 60 samples
Three clips from each of the twenty most common labels in
laion/vocal-bursts-gemini-segments
— 5,161 burst segments cut from real speech, labelled blind by Gemini 3.8 Flash.
What you are listening for. Each clip is one burst event and nothing else — the span
Gemini asserted inside a longer utterance, plus 50 ms. The label under each player is Gemini's
first choice; also lists its other guesses, because a burst is often genuinely two things
at once. The class in brackets is what our own corpus called the parent utterance — where
those two disagree is the interesting part, and disagreeing is the common case: the two agree on
the exact label only 19 % of the time.