Why AI-edited video sounds choppy (and the frame-snapping fix)
AI-edited video sounds choppy for three mechanical reasons: cuts placed off the frame grid force silent padding at every joint; audio butt-spliced at non-zero crossings clicks; and speech-model word timestamps run 50–100ms late, so cuts clip consonants. All three are fixable at the splice — frame-snapping, crossfades over room tone, and acoustic onset measurement.
By Marc Schultheiß — Founder of Never. Built the editing pipeline first — the frame-snapped splice, the acoustic cut refinement, the caption engine — then the company around it. · 2026-07-28
Bug one: the silence your concat filter adds
Cut a 29.97fps video at 4.63 seconds and there is no frame there. The renderer rounds the video to a frame boundary; the audio keeps its fractional length. Every segment now disagrees with itself by up to a frame, and the concatenation pads the difference — a few milliseconds of digital silence at every joint. One joint is inaudible. Forty joints are a rhythm of tiny dropouts, and the timeline stretches enough that word-timed captions land visibly late by the end. This is the bug most AI clip tools ship today, and it's why their output 'feels' edited even when the cuts are well-chosen.
Bug two: the click at every splice
Audio is a waveform; cut it anywhere but a zero crossing and the join step-changes — a click. The fix has existed since tape: run each segment's audio slightly past its video cut and crossfade the overlap with equal power over real room tone. Fading to silence instead punches an ambience hole your ear reads as 'someone edited here'. The J-cut is the same principle promoted to a technique.
Bug three: the 50-millisecond lie in every transcript
Word-level timestamps from Whisper-class models are systematically late at word starts and early at word ends — recognition lags the acoustic attack. Trust them and every cut-in clips a consonant. The fix is measurement: compute a 5ms RMS envelope around each boundary word, find the true onset, and place the cut outside it with a safety margin, clamped so it never crosses the neighbouring word.
None of these fixes is secret, and none needs a model — they need caring about joints at the sample level. That's the argument in practice for pipelines built by people who edited first: the failure modes are inaudible until you know them, and then they're everywhere. It's also why brand-locked output can sound hand-edited at scale for creators shipping daily.
Read the transcript
Why AI-edited video sounds choppy (and the frame-snapping fix) AI-edited video sounds choppy for three mechanical reasons: cuts placed off the frame grid force silent padding at every joint; audio butt-spliced at non-zero crossings clicks; and speech-model word timestamps run 50–100ms late, so cuts clip consonants. All three are fixable at the splice — frame-snapping, crossfades over room tone, and acoustic onset measurement.
Frequently asked questions
Why do cuts click in AI-edited clips?
The audio is cut mid-waveform and butt-joined: unless both sides happen to be at zero, the join is a discontinuity — a click. Editors avoid it with crossfades; software that concatenates segments skips them and ships the click.
What is frame-snapping?
Quantizing every cut boundary to the video's frame grid before rendering. Unsnapped cuts give each segment a video length up to a frame different from its audio; concatenation pads the gap with silence. Snapped, the rendered timeline equals the cut list exactly — which is also why captions stay in sync.
Why cut 40ms before the word timestamp?
Because speech models place word starts at recognition, not at the acoustic attack — consonant onsets begin tens of milliseconds earlier. Cutting at the timestamp clips the attack ('top' becomes 'op'). Measuring the onset on a short RMS envelope and cutting just outside it keeps the consonant.