Captions, transcripts and the room

Automatic captioning crossed a threshold in the last few years. For deaf and hard-of-hearing people, for anyone in a noisy place, and for anyone following a language they read better than they hear, it went from a novelty to something you can plan around.

It still fails, and it fails in a pattern worth knowing.

Where captions break

Names. Personal names, company names, product names. They are exactly what a meeting is about and exactly what a general model has least reason to recognise.

Technical vocabulary. Your industry’s terms, acronyms and jargon.

Crosstalk. Two people at once produces confident nonsense rather than a gap.

Accents. Accuracy varies substantially by accent, and the variation tracks how much of that accent was in the training data. This is not a small effect and it lands hardest on people already less well served.

Poor audio. A laptop microphone across a room, a bad connection, an echoing hall.

Notice how many are properties of the room, not the software. That is where the leverage is.

Arrange the room, not just the tool

The single most effective change is one person speaking at a time, close to a decent microphone. Everything else is secondary.

Practical things that work, and are reasonable to ask for:

Say names at the start. A round of introductions gives you a written list to correct the transcript against later.

Ask people to use headsets. A £20 headset outperforms any amount of processing.

Ask for an agenda and materials in advance. Knowing the topic makes garbled captions recoverable — you can reconstruct “the Q3 forecast” from a mangled line if you knew it was coming.

Ask for a pause between speakers. Half a second removes most crosstalk errors.

None of this is a special favour. It is the same request as “please share your screen”, and it makes meetings better for everyone — which is worth saying when you ask.

Transcript as well as captions

Live captions scroll away. Ask for a transcript too.

A transcript can be re-read at your pace, searched, corrected, and turned into actions with E4·L5’s structure — including its UNCLEAR section, which does double duty here: the things the transcript garbled land in it rather than being smoothed into confident wrong minutes.

Two cautions:

Consent first. E4·L5’s rule. Recording or transcribing needs everyone to know and agree, and it is a legal requirement in many places.

Correct the names before circulating. A transcript with everybody’s name misspelled is worse than no transcript, and fixing them takes a minute.

What captions do not give you

Honest limits, because expecting these and not getting them is worse than knowing.

Tone. Whether that was a joke, or annoyance, or hesitation.

Who is not speaking. Silence in a transcript looks the same whether the room agreed or went cold.

Overlap and interruption. The texture of a conversation flattens into sequential lines.

Professional captioners and sign language interpreters convey these, which is one reason automatic captions are an addition to that provision rather than a replacement for it. Where a meeting genuinely matters, ask for the human service — and note that in many jurisdictions employers and public bodies have a duty to provide it.

Try it now (6 minutes)

Turn on captions in your next call, and afterwards read the transcript.

Mark every place it got a name wrong. That list is the fastest possible measure of how much to trust it — and if you paste the list of correct names into the transcript request next time, many tools do noticeably better.

Check your understanding

1. Automatic captions fail most on: Long sentences Names, jargon, crosstalk, accents and poor audio Fast speech alone

2. The single most effective improvement to caption accuracy: A more advanced captioning service One person speaking at a time, close to a decent microphone Slower speech

3. What do automatic captions systematically fail to convey? The words themselves Tone, silence and interruption The order people spoke in

Recap

Captions became reliable enough to plan around and still fail on names, jargon, crosstalk, accents and bad audio. Fix the room first — one speaker, a real microphone, names read out, materials in advance. Take a transcript as well as captions, get consent, correct the names before circulating, and remember that tone, silence and interruption do not survive the transfer at all.