Tutorials Accessibility

Live Captioning for Streams and Events

Intermediate · ~18 min

Overview

Live captioning trades accuracy against latency against cost, and there is no option that wins on all three. Automatic captioning is instant and free and wrong in predictable ways. Human captioning is accurate and expensive and slightly behind. Hybrid approaches sit between. Choosing well depends on what the content is and who is relying on it. This guide covers the options, how to set each up, and the preparation that improves all of them.

What You Need

A clean audio feed. This determines everything downstream
A captioning source: platform automatic, a service, or a human captioner
A way to get captions into your stream, platform-native or an encoder input
Speaker names and specialist vocabulary prepared in advance
A test run before the event, not on the day
A fallback if the caption source fails mid-event

Steps

Fix the audio first, because everything depends on it

Automatic captioning accuracy is governed almost entirely by audio quality. Close microphones, one speaker at a time, minimal background noise, and no music under speech will improve results more than any change of captioning provider. A room where people talk over each other will defeat every option available.

Feed a vocabulary list where the system supports it

Most captioning services accept a custom dictionary of names, jargon, product names, and acronyms. Supplying it in advance dramatically reduces the errors that matter most, since proper nouns are both the highest-information words and the ones recognition reliably gets wrong.

Choose the approach against the stakes

For an internal update or a casual stream, automatic captions are a reasonable provision. For a public event, a legal or medical topic, an event where deaf attendees have registered, or anything where being wrong matters, human captioning is the appropriate choice. Deciding by budget alone is how organisations end up with unusable captions at exactly the events that needed them.

Set up the caption path and test it end to end

Captions reach viewers either through the platform's own captioning or by being injected into the stream at the encoder. Both need testing with your actual setup, not assumed. Check what viewers see on more than one device, since caption rendering varies considerably.

Account for the delay in how you run the event

Captions always lag speech, by a little with automatic systems and by a few seconds with human captioners. Presenters should avoid referring to things too quickly after showing them, and any interactive element (polls, Q&A cutoffs) should allow for people reading rather than hearing.

Publish corrected captions afterwards

Live captions are a best effort under time pressure and will contain errors. If the recording is published, replace them with a corrected caption file and a transcript. The live version served the live audience. It should not be the permanent record.

Pro Tips

Audio quality dominates. A better microphone improves captions more than a better captioning service.
Supply names and jargon in advance wherever the system accepts a custom dictionary.
Never put music under speech during a captioned live event. It degrades recognition sharply.
Ask presenters to speak one at a time and to pause between speakers. Overlap is what breaks live captioning.
Always replace live captions with corrected ones on the published recording.

Knowledge Base

What You'll Learn

Live captioning is a three-way trade-off, and the right answer depends on stakes rather than budget. Below: the options compared honestly, and why audio quality dominates everything.

The Three Options, Compared Honestly

Automatic captioning is instant, cheap or free, and available on most platforms. Accuracy is good with clean audio, a single clear speaker, and common vocabulary: and degrades sharply with accents, overlapping speech, background noise, and specialist terms. It also produces confidently wrong output rather than obvious gaps, which is its most dangerous property for content where precision matters.

Human live captioning, provided by a trained captioner working in real time, is substantially more accurate, handles overlapping speech and accents, and applies judgement about speaker identification and non-speech sound. It costs meaningfully more, needs booking in advance, and runs a few seconds behind.

Hybrid approaches, automatic recognition with a human correcting in real time, or a human respeaking audio into a trained recognition system, sit between the two on both accuracy and cost, and are increasingly common for medium-stakes events.

The decision should follow the stakes. If a deaf attendee has registered and is relying on the captions to participate, automatic captioning is not an adequate provision regardless of what the budget says.

Why Audio Quality Dominates Everything

Organisations reliably try to fix poor live captions by changing captioning provider, when the problem is almost always upstream.

Every captioning method (automatic, human, and hybrid) is working from the audio you supply. A recognition system cannot transcribe what it cannot distinguish, and a human captioner cannot type what they cannot hear. Both degrade with the same inputs, and both improve dramatically with clean audio.

The specific conditions that cause the most damage are consistent: overlapping speech, because two simultaneous voices are near-impossible to separate. Music or noise under speech, which masks exactly the frequencies speech recognition relies on. Distant microphones, which capture room reflections along with the voice. And unfamiliar proper nouns, which no amount of audio quality fixes but which a supplied vocabulary list does.

The practical consequence is that the money and effort that most improves live captioning is spent on microphones, on discipline about speaking one at a time, and on preparing a vocabulary list, not on the captioning service.

Where This Fits

This guide covers one specific part of captions and access. The wider picture, caption formats, reading speed, speaker identification, what automatic captioning still gets wrong, and a practical QA pass, is in Beyond Auto-Captions: Caption Quality, Styling, and Readability, which frames the discipline as a whole and links out to the detailed guides underneath it, including this one. If you are starting from scratch rather than solving a specific problem, read that first and come back here.

FAQ

Q: Are automatic live captions good enough?
A: For low-stakes internal content with clean audio and one clear speaker, often yes. For public events, specialist or precise subject matter, or any event where deaf attendees are relying on them to participate, no, automatic systems produce confidently wrong output rather than obvious gaps, which is worse than a visible failure.

Q: What is the biggest thing I can do to improve live captions?
A: Improve the audio. Close microphones, one speaker at a time, no music under speech, and minimal background noise will do more than changing captioning provider. Every method (automatic, human, and hybrid) works from the audio you supply and degrades with the same conditions.

Q: How far behind the speaker are live captions?
A: A short lag with automatic systems and a few seconds with human captioners, who need to hear a phrase before rendering it accurately. Plan around it: presenters should not reference things immediately after showing them, and interactive elements like poll cutoffs should allow time for people reading rather than hearing.

Q: Should I keep the live captions on the recording?
A: No. Live captions are a best effort under time pressure and will contain errors. Replace them with a corrected caption file and publish a transcript alongside. The live version served the live audience. It should not become the permanent record of what was said.