Trim dead air without a bad jump

Cut the breath. Keep the face legal. An authored jump cut on locked A-roll is pacing: same setup, dead air gone. A jump that lands on a blink, a swallowed consonant, or a teleport the model invented is a defect. Cover the authored skip with B-roll when the face would pop. Trim a video is the keep-range tool. It does not invent the in-between.

This is how to remove dead air from a talking-head you already signed. It is not the QC page for generate artefacts.

Name the skip: authored jump or artefact

Same look, opposite jobs. Write the name on the take before you cut.

Authored jump. You shot or generated a talking head. The camera is locked. You will delete the breath, the “um,” the false start. The face will pop forward in time. That is a cut you can point at on a timeline.

Artefact jump. One generate compressed three verbs. At 0:02 the bottle is closed; at 0:03 it is open and the hand never moved. There is no timeline cut. The file never contained the in-between. Trimming will not restore the pour. Regenerating that beat is the other page.

How to tell them apart:

Open the file. If there is only one clip and time skips inside it with no edit, it is an artefact. Stop. Do not “trim the weird bit” if that bit is the only proof.
If you can hear a breath or a stall you put there — or that a presenter left there — it is authored dead air. This page applies.
Scrub the skip in 1–2 frame steps. Did the subject travel. If not, artefact.
Would you have made this cut on a real camera day. “Delete the two seconds of staring” is a yes. “Delete the missing pour” is a no.

A cut to a new setup is not a jump cut. Hook face, then pack insert, is a cut. Do not treat that join as a problem this tool should soften.

Duration artefacts hide in long single-pass generates. This page assumes the talking-head is continuous enough that the only thing you are removing is idle. If the mouth skips a syllable inside a word, that is not dead air. Leave it or replace the take.

Mark the breaths, then keep one range at a time

cut_video with one [start, end) is a straight trim. Several keep-ranges concatenated is how a flub in the middle disappears. Both are legal. Both need marks.

Procedure:

Watch at 1× with sound. Write every stall: 0:04–0:06 breath, 0:11 “so yeah”, 0:18 look-down.
Watch muted. Confirm the stall is idle, not a reaction you need.
For each mark, decide jump or cover:
- Jump: locked camera, eyes open, no blink on the join, no half-word.
- Cover: the join would land on a blink, a chewed consonant, or a pop so large the face feels like a glitch.
If the dead air is only handles at head and tail, one keep-range is enough. Trim a video that range. Do not over-cut into the first word.
If the dead air is in the middle, list keep-ranges in order and concatenate. That is the same tool with more segments, not a second app.
Keep the parent. A trim is a child. Handles on the parent are how you undo a greedy cut.

Checkable join test, per cut:

Play three frames before and three frames after.
Eyes: both open, or both in a matching blink. A closed-to-open pop is the bad jump.
Mouth: not mid-phoneme. Cut on silence or on a finished word.
Head position: a large translation in one frame is the pop. If the presenter leaned in during the breath, cover it.

Do not trim a take you have not signed. A shorter miss is still a miss. Picture lock first, then the keep-range.

Frame-level honesty depends on the source frame rate. Do not expect a 24 fps generate to give you 1/100s precision. If the only legal cut is one frame into a word, you do not have a legal cut. Cover or recut the sentence.

Cover the blinks with B-roll

When the face would pop, do not “live with the energy.” Cut away.

Procedure:

On each join you marked cover, note the duration of the skip you are hiding. Usually 8–20 frames of B-roll is enough. You are covering a splice, not replacing the scene.
Choose cover that is already true: the locked pack, the hands, the result, a filmed insert. Image-to-video from the approved still is allowed. A random stock wink is not.
Cut to the cover on the splice, not a beat later. Late cover advertises the pop.
Come back to the face on a finished word, eyes open.
Do not park a caption, emoji, or whoosh on the splice and call it covered. Mute viewers still see the face jump. Type is not frames.
If you have no honest B-roll, generate a two-second pack insert from the locked still and use that. Empty cover is worse than a small authored jump on a clean blink-free join.

Allowed covers:

Product insert from the same SKU still.
Hands doing the verb the sentence names.
A result plate (stain, readout, before/after) you already approved.

Forbidden covers:

A caption that says “watch this” over the pop.
Zooming in so the jump is a different crop of the same glitch.
Slow motion on either side of the splice.
A second talking-head take with a different shirt as “B-roll.”

The glossary rule is the whole craft: jump on A-roll to kill dead air; cut to B-roll when the face would pop on a blink.

If every splice needs cover, your A-roll is too loose. Re-record or regenerate a tighter take. Cover is for a few splices, not for building the edit out of wallpaper.

Recaption and retime after the trim

Captions follow the child, not the parent. A burn on the untrimmed file will not travel.

Finish order:

Lock the trimmed picture. Watch the whole child at 1×. Count remaining breaths. One leftover stall is allowed if it is a thought. Three is a trim you did not finish.
Listen. Native audio that still contains a cut breath in the waveform means you trimmed picture and left a handle in sound, or the other way around. Picture and sound should share the keep-ranges unless you are about to replace audio.
Burn captions on the child. Add captions to video last. 165 language codes. BASIC at 1× unless you have a reason for DYNAMIC at 2×.
Check caption timing on the first and last thirty seconds. A constant offset is what you want after a head trim. Progressive drift is a frame-rate mismatch; an offset will not fix it.
Recheck safe area. A recrop after trim is a new frame.

If you already had an SRT on the parent, shift it only when the trim was a single head or tail chop by a constant. Middle keep-ranges mean a new caption pass. Do not hand-nudge forty cues to follow three splices.

The AI video generator produced the parent. Trim is a child. Archive both.

Cut the breaths, cover the blinks, recaption the child in Versely. Bring the keep-ranges. Do not bring “make it snappier.”

FAQ

Is every hard cut on a talking head a bad jump?

No. Same shot, idle removed, eyes open, word finished: that is authored pacing. A pop on a blink, or a teleport inside one generate, is the defect.

Can I hide a generate artefact the same way I hide a breath?

No. B-roll can cover a splice you made. It cannot restore a pour the model skipped. Regenerate that beat.

Should I caption first so I can see the words while I trim?

No. A later trim is a recaption. Trim the keeper, then burn.