Updated October 2026 · Speech recognition and on-device transcription
Live captions turn speech into text as you talk, using real-time recognition. They help viewers who are deaf or hard of hearing, and viewers watching without sound. Expect a short delay and some errors, check the settings before going live, and keep the captions readable against the video.
Live captions have moved from a niche feature to a common option on video platforms and streaming tools. A caption appears under the picture as the speaker talks, which helps viewers who are deaf or hard of hearing, viewers in a noisy place and anyone watching with the sound off. For a creator, captions also improve accessibility and can keep people watching longer. They come with trade-offs that are worth understanding before going live.
This post explains how live captions work, the delay and error trade-offs, the settings to check and the habits that make captions useful rather than distracting.
Live captioning runs a speech recognition model on a continuous stream of audio. The model produces partial text as the speaker goes, then revises it as more context arrives. That is why a caption can change a word after it first appears. The process is fast, but it cannot wait for a full sentence without adding delay, so it makes a judgement with less context than a recording would allow.
The result is a good first pass, not a finished transcript. Expect the captions to be less accurate than a transcript made afterwards from a clean recording.
Live captions usually lag the speech by a short time. The lag depends on the tool, the connection and the length of phrase the model waits for. A viewer may see the caption a moment after the words are spoken, which can feel slightly out of step. For fast-paced talk, the lag is more noticeable, and the caption may seem to arrive in bursts.
Test the delay before you go live, by speaking a few sentences and watching the caption. If the lag is large, consider slowing your pace slightly or choosing a different setting.
Live recognition makes errors with names, technical terms, numbers and accents. A wrong word can change the meaning of a sentence, and a viewer may notice the error more than the correct text around it. Use a shared glossary of names and terms before the stream, and check the most important words in the first minutes to see how the tool handles them.
Consider telling viewers that captions are automatic and may contain errors. A short note sets fair expectations, and it helps viewers who rely on the captions to understand the limits.
Captions should be large enough to read on a phone, with enough contrast against the picture. Keep each line short, and avoid changing the position of the captions during the stream. A caption that jumps around distracts the viewer from the speaker. The guidance in the readability post in this series explains the basics.
Captions depend on clean audio, and a noisy microphone produces more errors. A steady, close microphone, a quiet room and a level that is not clipping all improve the captions as well as the sound. A live stream with good audio gives the recognition a better chance, so treat the audio setup as part of the captioning.
If you save the stream, keep the captions as a file. A saved transcript can be corrected after the event, and the corrected captions can be attached to the recording. Editing the captions afterwards makes the replay more useful, because the errors that mattered in the live version can be fixed for later viewers.
A live caption feed can fail in the middle of a stream, through a dropped connection, a crashed app or a microphone that stops sending audio. Decide in advance what you will do: whether the stream continues without captions, whether a second caption source is ready, and how you will tell viewers. A short pinned message that explains the problem and the expected return is far better than silence.
Rehearse the recovery once. Turn the captions off and on again during a test, and time how long it takes to restore them. A rehearsed fix turns a panic into a routine change.
Live captions are a first pass, and they cannot match a carefully checked transcript. Features and accuracy vary between platforms and tools, and claims from vendors are not independent measures. Test the setup before a live event, and be ready to explain that the captions are automatic.
Good captions begin with a good microphone signal. The Clean Streaming Mic offers live noise cancellation for OBS, Discord and Zoom, which helps keep the speech clear before it reaches the recognition. The Tunedzy Go page covers cleaning a phone recording. For the wider picture of recognition, read the on-device transcription post.
Record voice-overs with live noise removal, or clean an existing file. Runs in your browser, nothing is uploaded.
They are a good first pass, but they make errors, especially with names and terms. Tell viewers they are automatic.
The recognition waits for context before it finalises text. Some lag is normal.
Often yes, depending on the tool. A saved file can be corrected and attached to the replay.
Yes. They help anyone watching with sound off, in a noisy place or in another language.
