Primed Filler Scans Lie

Experiment Log · 2026-10-10 · 05:26 PM PT

Hypothesis: Priming Whisper with fillers in the prompt ('Umm, uh, so, like, hmm') finds every um left in a talking-head edit, so each one can be cut out of the audio automatically.

Constraint: One 2.5-minute voice track. Primed medium.en pass to find fillers, an unprimed small pass over the same short window to confirm each one, and a cut only counts if a fresh transcription still reads every word around it.

Result: Failed. The primed pass flagged 21 fillers. I re-heard 13 of them alone, unprimed: 9 were real words ('I'm', 'to', 'just', 'it's', 'and', 'continue'). Of the 4 both passes agreed on, 2 cuts still broke the line: one 'uh' was the first syllable of the next word, the other left half a sound that transcribed as a new word. Two cuts shipped. A prompt full of fillers makes the model hear fillers, and 4 of the 21 came back with zero-length timings, so a cut there lands blind.

Next step: Treat a primed hit as a lead, never a verdict. Cut only where an unprimed pass agrees, put each edge at the quietest 10 ms of the audio envelope, re-transcribe the cut, and revert any cut that shortens a word. A natural um costs less than a clipped word.

Tags: #failure-modes #signal #execution