Tutorials Accessibility

Live Captioning for Broadcast: Speech Recognition, Human Co-Pilots and Encoders

By Infinite Creation · Professional · ~15 min

A 1960s mid-century comic book style editorial illustration showing two people in a broadcast gallery, one holding a printed sheet and the other watching a monitor with a presenter and a band of wavy lines beneath, with a stopwatch beside a second screen.

Overview

Live captions for broadcast used to be typed by trained captioners. Makers now sell speech recognition that produces the first caption stream, with a person added or not. This guide is for broadcast engineers, captioning coordinators and producers who have to choose or check such a service. It sets out what Verbit and AI-Media state, which parts they leave without a figure, and a test you can run on your own programmes. We did not test any tool on this page. Each statement about a tool comes from the maker's own page as read on 8 October 2026, or from a trade source named in the Sources section.

You will need:

Three to five recordings of your own live programmes, including one with crowd noise or several talkers
A list of the names, places and terms that appear in those programmes
A reference transcript for each, produced by a person, so you can count errors
Your regulator's or buyer's caption quality requirements, read in full

Steps

Decide whether you need captions from speech recognition, a person, or both

Verbit's 2022 launch of its broadcast product, covered by Sports Video Group, describes an automatic captioning product offered in two forms, ASR alone and hybrid. In the hybrid form a client can add a "human co-pilot" who controls the live session, covering corrections, caption placement and subject changes. The article says specialists prepare and review the content loaded into the engine before each session. These are the maker's claims. Decide first which of those two arrangements your programmes can accept.

Read what Verbit's own page states today

Verbit's home page offers real-time captions powered by Captivate for meetings, depositions, broadcasts and classrooms. It calls Captivate customizable speech recognition and says the service is backed by AI plus human expertise. It does not say that live captioners are human, and it gives no accuracy figure for live captioning. The article from 2022 also gives only descriptions such as "highly accurate", with no numbers.

Read what AI-Media states about its encoders

AI-Media's own release page returned an error when we tried it, so this rests on Sports Video Group's report of April 2026. It says the LEXI Text Encoder and LEXI Voice Encoder are the maker's first new encoder hardware in more than a decade, with 4K over 12G-SDI, support for SDI, IP and cloud workflows, and LEXI Live Sync, which gives configurable timing alignment for captions and translations. It also reports a Hardware-as-a-Subscription model. The article names no speech engine, no languages, no price and no shipping date.

Load your names and terms first

Whatever the service, give it the names of presenters, teams, places and products before the broadcast, and ask how it takes them. Verbit describes pre-session preparation. Ask any other vendor the same question and ask what happens to a name it was not given.

Run your own recordings and count errors

Play each recording through the service in the way it will run at air time. Compare the captions with your reference transcript. Count wrong words, missing words and wrong names separately, because a missing sentence and a wrong name harm the viewer differently. Note the delay between speech and caption on a stopwatch, and note what happens when two people talk at once.

Plan the fallback before you go live

Decide who watches the captions during the broadcast, what level of error makes them take over or switch to a person, and how the encoder behaves if the connection to a cloud service drops. A written fallback matters more than the average score.

Pro Tips

Test on your worst audio, not your best. Crowd noise and crosstalk are where captions fail.
A tool that is "highly accurate" in a press release has told you nothing you can contract on. Ask for a measured figure on your material.
Keep the reference transcript. You will want it again when the vendor updates the engine.

Knowledge Base

Going Deeper

The question is not whether speech recognition can caption live speech. It is what happens on the minute it gets a name wrong.

Why the figure matters more than the adjective

Both pages we read describe their output in words such as accurate and customizable. Neither gives a measured rate, a list of conditions where the service struggles, or a delay. Without those, you cannot compare two services, or the same service before and after an update. Measure on your own material, and write the method down so the next test is comparable.

Encoders and engines are separate decisions

An encoder inserts captions into the signal. A speech engine writes them. AI-Media's release describes new encoder hardware, and it names no engine. Verbit describes an engine. Buying one does not answer the question about the other, so ask each vendor which part it supplies, and which part it takes from a partner.

What we did not read

The regulator's own caption quality rules were not reachable when we tried. This guide therefore says nothing about what a regulator requires for live captions. Read your regulator's text, and your buyer's specification, before you rely on any service.

Where This Fits

A speech engine writes live captions with no second pass, so the quality and timing rules in Beyond Auto-Captions are what its output is measured against.

FAQ

Q: Can speech recognition replace human live captioners?
A: We did not test it and the makers' pages do not settle it. Verbit offers a hybrid with a human co-pilot, which suggests the maker sees a role for a person. Test on your own programmes.

Q: Which service is the most accurate?
A: We cannot say. The pages we read give no accuracy figure. Measure on recordings of your own programmes.

Q: Does a captioning encoder include the speech engine?
A: Not necessarily. AI-Media's release describes encoders and names no engine. Ask each vendor what is included.

Vocabulary

ASR Automatic speech recognition. Software that turns speech into text without a person typing each word. Live captioning uses it to produce a first caption stream that a person may then watch, correct or replace.

Human co-pilot Verbit's term for a person who controls a live speech recognition session, correcting errors, placing captions on screen and switching subject. The term is the maker's own, and a client may choose to add or leave out this person.

Caption encoder A device or service that inserts caption data into the broadcast signal so that receivers can show it. It is a separate part from the speech engine that writes the words, and a buyer may purchase one without the other.

12G-SDI A single-cable connection that carries 4K video between devices in a broadcast facility. AI-Media states that its new caption encoders support 4K over it, so they can sit in a 4K signal path without extra conversion.

Sources

Two trade articles and one maker's page were read on 8 October 2026. Nothing was tested. Checked 8 October 2026.

Sports Video Group, Verbit launches human-powered ASR captioning (12 October 2022). A trade article on the launch. Describes ASR and hybrid versions, a human co-pilot option, pre-session preparation and no numeric accuracy.
Verbit, home page. Offers real-time captions powered by Captivate for broadcasts among other uses. States no accuracy figure and does not say live captioners are human.
Sports Video Group, AI-Media launches LEXI Text Encoder and LEXI Voice Encoder (15 April 2026). A trade article. Lists 4K over 12G-SDI, Live Sync, SDI, IP and cloud workflows and a subscription hardware model. Names no engine, languages, price or date.

AI-Media's own release page returned HTTP 429 and its product pages were not reached, so every AI-Media statement rests on one trade article. The Verbit launch article is from 2022 and the product may have changed. The regulator's caption quality rules were unreachable and are not summarised here.