Speaker IDs on two-person UGC

When both people talk, write speaker IDs. Auto captions transcribe speech. They will not reliably invent [Maya] and [Jordan], and they will collapse overlap into one cue. SDH is the caption flavour that includes who is speaking and the non-speech beats a mute viewer still needs. One-host UGC often needs zero IDs. A two-hander that ships speech-only is dialogue with no cast.

This is the two-person UGC procedure. It is not a full DCMP style-guide reprint.

Hear two mouths before you caption as one stream

Two-person UGC is any file where a mute viewer cannot tell who is talking from the words alone: a dual talking-head, a founder plus a customer, a voice-off producer, a phone call, a “wait, you actually use this?” cut.

Procedure before captions:

Watch muted. If you cannot assign each line to a body, you owe IDs (or a crop that makes the speaker obvious).
Watch with sound. Count talkers. num_speakers on a dub later will need the same count. Caption IDs need names or roles now.
Write a cast list using the names the ad already uses. If the ad never names them, use short roles: Host, Customer, VO. Do not invent Woman 2 if Friend is clearer. Stay consistent for the whole file.
Mark overlap timestamps. Overlap is both an ID problem and a timing problem. You will split those cues, not let the engine interleave two sentences into one.
Decide whether this file is speech-only captions or SDH. If a door slam or a laugh is the cut, you need the bracket. If the only extra is who is talking, IDs without a full effects list can still be enough for a short ad.

One talking head plus a music bed is not two-person UGC. Do not ID the beat drop.

A PiP face over a pack, with only the face speaking, is one speaker. Do not ID the bottle.

The AI caption generator will happily emit one stream. Your pass is to split it.

Add IDs where the picture fails, not on every line

A speaker ID is a prefix in the caption track, timed to the line: [Maya], Host:. It is not a lower-third graphic. You may also have a name plate. The mute reader of the burn still needs the prefix when the plate is gone or both faces are in frame.

When to ID:

The speaker is off screen.
A new voice enters.
Both mouths are in a two-shot and your burn is centered (placement cannot do the job).
Overlap.

When not to ID:

A single-host read with no other voice.
A stretch where the same person is clearly the only face, after you already labelled the switch.
Every line of a two-hander once the pattern is established — re-ID after a gap, a new scene, or a third voice.

Procedure:

Transcribe or burn a first pass on the locked two-person file. Add captions to video gets you speech. 165 language codes.
Open the cues. Split any cue that contains two voices. One voice per cue.
Put a short ID on the first cue of each stretch that would be ambiguous muted. Keep it short. A long role title blows caption line length on 9:16.
Drop the ID once the speaker is visually obvious for a stretch, unless you are delivering full SDH to a spec that wants names throughout. Most UGC ads are not that spec. Ambiguity is the test.
Off-screen VO: ID and, if your look allows italics, italicize. A disembodied line that looks like the on-camera person is a lie.
Karaoke / DYNAMIC: put the ID on the first token of a new speaker or the accent walks across a line whose owner is unknown. Word-highlight does not replace the label.

Do not stack two names on one block. Split.

Do not use the ID as the hook. [Maya] oh my god still has to be readable as a claim. If the first second is only a name tag, you do not have a first-second hook.

If you generated the two voices with multi-speaker TTS, keep those voice names aligned with the caption IDs. A wav labelled “Speaker A” and a burn labelled “Host” is homework for the viewer.

SDH extras only when the sound is the beat

SDH assumes you cannot hear the mix. It adds [door], [music], [laugh], speaker IDs. Auto-transcribe will not invent those brackets. If the brief is SDH, someone authors the non-speech cues.

For two-person UGC ads, be stingy:

If a sound causes the cut or the joke, write it. A mute viewer will not infer the knock that makes them both look left.
If the sound is a bed, skip it. A wall of [upbeat music] on a 15-second ad is a second script.
Speaker IDs are the SDH piece you almost always owe on a two-hander. Non-speech is case by case.
Do not label a file SDH because you added [Maya] once. SDH is speech plus non-speech plus IDs as needed. Speech-only with occasional names is “captions with IDs.” Call it honestly.

WCAG 2.2 Success Criterion 1.2.2 wants captions for prerecorded synchronized media, including non-speech information when it matters. A six-second pack ad that is already designed as mute-readable type is a different shape than a two-minute dialogue. Still: if both people talk, IDs are the minimum so the mute viewer has a cast.

Forced narrative is not this job. FN is the few lines a same-language viewer still needs (a foreign sign, a third language). Two English talkers getting English IDs is SDH/captions, not FN.

Burn, then watch mute as a stranger

The QC is a person who cannot hear and does not know your friends.

Checklist:

Picture lock on the two-person cut. A later trim is a recaption of every ID.
Burn BASIC (21 looks, 1×) unless you have a reason for DYNAMIC (9 looks, 2×). Two-hander plus karaoke plus IDs is a lot of motion. Readable blocks take IDs more calmly.
Place in the safe area. IDs add a line. That line still cannot sit under the username or the reply bar.
Play muted. At each switch, can you tell who spoke. If you use faces as the only ID, cover the screen with your hand except the caption row and try again — that is how a lot of people actually watch.
Play overlap moments. Two names, two cues, no mashed sentence.
Confirm the SKU and the offer are still the point. A perfectly labelled two-hander that never shows the pack is a podcast clip.

If you will dub this file, IDs belong in the target-language captions after the dub. Do not translate [Maya] into something else. Names stay. Roles may translate.

Archive the cast list with the cut. Take twelve of the same two creators should not rename them.

Label the two-hander, then burn in Versely. Bring the locked file and two names. Do not bring a single-stream transcript and call it SDH.

FAQ

Do I need speaker IDs on a one-person UGC ad?

Not until a second voice enters (a VO, a phone, another avatar). ID the first line of the new voice, not every line before it.

Is a lower-third enough instead of caption IDs?

No. A lower-third is a graphic. Mute autoplay readers of the burn still need the prefix when both talk or the speaker is off screen.

Will auto captions add [laugh] and names for me?

No. They transcribe speech. IDs and non-speech are a pass you author after that.