Words That Arrive Too Late: Putting Live Captions on Council Webcasts to the Test
Photo: Shixart1985, CC BY 2.0, via Wikimedia Commons
When a council streams a planning committee or a full council meeting, the recording is usually treated as the durable artefact: it is indexed, archived, and occasionally given a transcript weeks later. The live stream, by contrast, is treated as a convenience. That assumption deserves challenge. For a resident who is deaf or hard of hearing, and who wants to follow a licensing decision as it is being made, the live caption is not a convenience. It is the only route into the room.
Pre-recorded transcripts are now audited with some regularity. Live captions largely escape that discipline, for three reasons: they are ephemeral, they carry no obvious failure signal for the people who do not need them, and vendors report accuracy figures that are difficult to verify. The result is a gap between what an organisation believes it provides and what a viewer actually receives.
What the standard asks, and what it does not
Success Criterion 1.2.4 (Captions (Live)) sits at Level AA in WCAG 2.2 and requires captions for all live audio content in synchronised media. Under the Public Sector Bodies (Websites and Mobile Applications) Accessibility Regulations 2018, public sector bodies in the UK are expected to meet WCAG 2.2 AA, with the technical detail mapped through EN 301 549. The Equality Act 2010 adds a separate reasonable adjustments duty that does not turn on the regulations at all.
The important omission is that neither WCAG nor EN 301 549 sets a numeric accuracy threshold, and neither prescribes human captioners over automatic speech recognition. The mechanism is not the test. Accuracy is. That means an organisation has to decide, in writing, what standard it is holding itself to — and then measure against it.
In practice, many accessibility practitioners work to 98–99% accuracy for pre-recorded captions and accept a lower figure live. Whatever figure a council adopts, it should be published in the accessibility statement alongside a latency target, so that residents know what to expect and web teams know what to defend.
The audit routine: five passes
The routine below can be run against three meetings — ideally one full council, one regulatory committee, and one meeting with substantial public participation — and completed by two people in roughly two days.
Pass one: the platform, before anyone speaks
Join a test stream or a low-stakes briefing and check the basics. Are captions on by default, or hidden behind a menu that a first-time viewer will never find? Can the viewer change font size, background opacity, and position? White text on a transparent plate will fail against a bright chamber wall, so check the caption styling against contrast requirements rather than against a developer's monitor. Test on a council-issued laptop, a personal phone, and — if the authority uses one — a shared device in a library. Confirm that captions survive full-screen mode, picture-in-picture, and the mobile player, where they most often disappear.
Pass two: latency
Latency is the measure most closely tied to comprehension, and the easiest to record badly. Do not guess. Play the stream on a laptop, keep a stopwatch on a phone, and note the moment a distinguishable phrase begins and the moment its caption completes. Take ten samples across the meeting, including at least three from public speakers, and report the median and the worst case.
A useful working target is a lag of no more than four seconds for public question time. Human captioners typically land within two to three seconds; automatic recognition on a busy audio feed often drifts to six or more, and drifts further as the meeting goes on and audio quality changes. Where a question from a resident is captioned eight seconds after it is asked, the answer has already begun before the viewer has read the question.
Pass three: accuracy, measured by category
Aggregate word error rate flatters live captions, because most speech in a council meeting is predictable. The failures that matter cluster in four places, and they should be counted separately: proper nouns (councillor surnames, ward names, site addresses), figures and currency, acronyms, and negation. A dropped 'not', a mangled '£4.2 million', or a requisition reference rendered as letters rather than numerals can invert the meaning of a decision.
Take a continuous five-minute stretch that includes a public speaker with a regional accent or a dialect different from the chair's, transcribe it verbatim, align it against the captions, and count substitutions, insertions, and deletions against the total word count. Report the category breakdown alongside the headline figure. If a supplier claims 95% accuracy, the useful question is not whether that number is true in aggregate but what the error rate was on the nine words that carried the decision.
Pass four: speaker identification
In a chamber, attribution is half the meaning. Check whether captions name the speaker, whether the name is correct, and whether the mapping survives overlapping voices, interventions from the chair, and remote participants joining a hybrid meeting. A caption stream that alternates between 'Speaker 1' and 'Speaker 2' is technically present and practically useless when four members are speaking from the same bench. Where automatic recognition cannot attribute reliably, a named human captioner working from the agenda and the seating plan will.
Pass five: what happens when it breaks
Every live caption service fails eventually, and the audit should find out how. Disconnect the captioner mid-stream, or ask the supplier what happens when recognition confidence drops. Is there a visible on-screen message, or does the text simply stop? Is there a published telephone number for the meeting? Is British Sign Language interpretation available, and is it clearly distinct from captioning rather than offered as a substitute? After the meeting, check the archive: are captions retained on the recording, are they correctable, and how quickly is a transcript published? Treat the archive as a publication in its own right, not as a by-product.
Recording the results
A simple table keeps the audit defensible and makes regression testing straightforward next year.
| Criterion | Measure | Working threshold | Result |
|---|---|---|---|
| Latency | Median lag across 10 samples | ≤ 4 seconds | |
| Accuracy | Word error rate, five-minute sample | Organisation's published figure | |
| Proper nouns | Errors per 100 nouns | ≤ 2 | |
| Speaker attribution | Correct naming rate | ≥ 95% | |
| Failure state | Visible message within 10 seconds | Present | |
| Archive | Captions retained and correctable | Yes |
What good looks like
Few authorities can justify a human captioner for every internal briefing, and it would be disingenuous to pretend otherwise. A proportionate position is to reserve live human captioning for full council, regulatory committees, and any session with public participation, while permitting automatic recognition for low-stakes officer briefings that are post-edited before publication. The point is that the choice should be deliberate, documented, and matched by a published accuracy and latency commitment, with a named owner and a fallback plan that a viewer can actually find.
None of this is exotic. It is the same discipline already applied to transcripts, applied to the moment that matters most. A resident following a licensing application at seven o'clock on a Tuesday evening cannot wait for Friday's transcript. The caption either arrives with the words or it does not.
