Captions on talking-head vs B-roll

Talking-head (A-roll) has speech; caption that speech. Silent B-roll has no speech; do not auto-caption it. Write an overlay if the mute viewer still needs a claim. Mixing the two jobs is how a product plate ships with empty cues, or a face ships with a headline the mouth never said.

The tools are already split. Add captions to a video transcribes audible speech in 165 language codes and burns a look. Add a text overlay burns a line you authored. Text overlay is copy nobody had to say. Auto captions are a timed transcript. A cut that switches from face to insert needs both, applied to the right ranges, not one pass on the timeline hoping the generator "knows."

A-roll has speech

A-roll is the spine: a person talking, an interview, a founder to camera, a UGC creator holding the product and saying the script. The mute feed still has to read that script. Burned-in speech captions are the job.

Place them off the mouth. Centre-bottom is where talking-head models and webcams put the face. Classic, Paper, and Ink default to the bottom; Classic Top / Paper Top / Ink Top exist because of this collision. If the chin is the shot, raise the block. Karaoke (Snap, Reveal) punches on the lower third and will scale a word on the teeth. Use timed families on A-roll only when the words are the performance and the face has room.

Read names and numbers before you burn. A talking-head is where those strings live. Accents, a cheap lavalier, a kitchen extractor — accuracy drops. The captioner will not know the product spelling. You do.

A-roll can be the whole ad. UGC often is. Then you may still want one authored hook on top ("3 ingredients, no foam") because the spoken first sentence was a greeting. That hook is overlay. The rest is transcript. Two layers on one face: hook at the top, speech mid-lower, mouth visible.

Do not generate a host to "give the B-roll someone to caption." If the spine is a SKU, the SKU is the A-roll in editorial terms and it still does not have speech. Inventing a presenter so the captioner has audio is a different video.

B-roll is often silent

B-roll is the insert: product on a table, process, street, the thing the sentence is about. Good B-roll has no native dialogue. The glossary is specific: no hard cuts inside the clip, motion slower than the speech, nothing more interesting than the line being read.

If you feed a silent insert to auto captions, the engine has nothing honest to hear. You spend a 1× or 2× caption render on room tone. You may get stray words from a TV in the next room or from music. You will not get the claim. That is the failure mode the VEED task already names: a clip with no audible words spends the row on nothing.

When B-roll plays under a voiceover, the speech is on the A-roll track, not in the insert file. Caption the mixed edit (voice + picture), not the raw B-roll clip. If you captioned the interview and then covered it with inserts, the captions can stay if they were burned on the mixed timeline. If you captioned the talking-head file and then replaced the picture with B-roll in another tool, you now have speech captions on a picture that has no mouth — which is fine — timed to a soundtrack you may have trimmed. Confirm the clock.

When B-roll is the whole post — faceless product, ASMR pour, a 6-second pack shot — there is no transcript. The mute viewer needs a line you write. Overlay. One claim, large, off the product and off the platform chrome. Headline or a top overlay. Not Snap. Snap has no word to time.

Music-only B-roll is still silent for ASR. Do not transcribe a beat.

Overlay vs auto-captions

Use this test on every range of the cut:

Did a person say this line on this soundtrack? Yes → auto captions. No → overlay, or nothing.

"Nothing" is allowed. A 1-second insert of the product does not need type if the surrounding A-roll captions already named it. Type on every shot is how a 20-second ad becomes a slide deck.

When A-roll and B-roll alternate in one file, one caption pass on the mixed export is usually enough: the transcriber follows the voice, including over the inserts. Cues will sit on the product for a second. Check that they do not cover the label on the pack. Raise or split those cues. You are not switching tools mid-file; you are placing.

When the insert needs a different sentence than the voice ("Clinically tested" while the founder talks about texture), that sentence is overlay. If you try to cram it into the caption track, you are mixing authored copy into a transcript. The next recaption will wipe it. Keep authored claims in the overlay tool so a recaption of the speech does not delete the legal line.

Timed overlays are for a sequence of authored claims. Speech captions are for speech. A listicle that is only type over B-roll is a timed-overlay video. Running add_veed_captions on it because "it needs captions" is the silent-file bug.

Credits: speech burns are 1× basic / 2× dynamic per file. Overlays are a separate editor render. Captioning a silent B-roll plate to "be safe" is a 1× spend with nothing to show. Skip it.

Cutting between them

A practical assembly:

1. Lock the mixed cut: A-roll, B-roll, voice.
2. Caption the mixed file in the spoken language. Basic family. Preview the look on a few seconds.
3. Watch on mute. On A-roll shots, is the mouth clear? On B-roll shots, is the pack label clear? Move the block, do not shrink it into the chrome.
4. Add the hook overlay if the first spoken line is not the claim.
5. Recaption only if you changed the soundtrack or the in/out.

Reframing A-roll to 9:16 after captioning will throw placement. Crop, then caption. B-roll generated at 9:16 from the start is cheaper than a widescreen insert you later punch.

Burned-in captions on a talking-head are how silent autoplay still has a person. On B-roll they are optional type. The AI caption generator is the speech door. The overlay task is the claim door. Faceless channels live on the second. Creator UGC lives on the first. Most ads are both, in one file, with two tools.

If the insert is more interesting than the sentence, that is a B-roll brief failure, not a caption failure. Type will not save a cutaway that upstages the line.

Caption the spine in Versely.

FAQ

Should I auto-caption B-roll that has a little camera audio?

If the audio is the voiceover you want read, yes — on the mixed file. If the audio is handling noise and a café, no. You will caption the café.

Can Snap captions run over product inserts?

Only while someone is still speaking on the soundtrack. Silent pack shots have no tokens to pop. Use an overlay for the claim.

Do I recaption after I drop in new B-roll?

If the voice and duration did not change, placement is the check, not a new transcript. If you trimmed to fit the insert, recaption. Burns do not ripple.