Why filler cuts sound choppy
Most choppy cuts have one of three causes: the cut clips the start or end of a neighbouring word, the join clicks because the waveform jumps, or every pause is closed so the speech sounds breathless. Filler tools acknowledge the first one: Descript has an “Avoid harsh cuts” option that skips fillers it can’t remove cleanly, and Riverside groups risky cuts under “Harsh cuts.”
Fix the sound: where and how to cut
Cut where one sound ends and the next begins. Speech-recognition timestamps are approximate, so a cut placed on them can shave a consonant. A forced aligner, which matches each word’s sounds to the waveform, gives tighter boundaries.
Crossfade each join over a few milliseconds so the waveform doesn’t jump and click. Keep a short natural pause where the speaker was thinking; remove the “um,” not the breath.
Hide the jump in the picture
- Cut away to B-roll over the join: the product, the screen, the place being described.
- Punch in or out slightly at the cut so the jump reads as a deliberate camera change.
- Put a caption, title or graphic on screen across the join.
- With two cameras, switch angle at the cut.
Don’t over-cut
Removing every filler and every gap can make a natural speaker sound like a read-out. Cut the hesitations and tics that slow the point down, keep the pauses that give it rhythm, and listen to the whole clip once at normal speed before you publish.
How Shorty handles it
When Shorty removes fillers, it cuts the original recording at 48 kHz on phonetic boundaries from a forced aligner, with a 4-millisecond equal-power crossfade at each join. It refuses any cut next to a word it can’t place with confidence, never cuts through a word it didn’t choose, keeps pauses that aren’t part of a cut, and lands picture cuts on whole frames so lips stay in sync. Ask for B-roll, punch-ins or captions over the cuts in the same brief.
See it on four real talks
We ran four public talks through the same tool the director uses inside an edit. Across them, 107 words came out in 80 cuts, including every “um” and “uh” in the transcripts, and every kept word survived in order. The launch post plays each one before and after, with every removed word struck through.
| Recording | Length before → after | Words cut | Cuts | Ums and uhs left |
|---|---|---|---|---|
| Nik Storonsky, Revolut | 2:00 → 1:49.8 | 40 | 28 | 0 of 13 |
| Prime Intellect, NVIDIA event | 1:00 → 0:53.2 | 38 | 27 | 0 of 16 |
| Arjun Karanam, Trajectory | 2:00 → 1:54.4 | 22 | 21 | 0 of 8 |
| Elon Musk, Tesla Roadster event | 2:00 → 1:58.4 | 7 | 4 | 0 of 1 |
Sources: Shorty News: four talks before and after. Measured on excerpts of public YouTube talks on 20 October 2026. Results depend on the recording.