Captions and Subtitles: Formats, Rules and Tools

Captions are timed text for people who cannot hear the audio, so they name the speakers and describe sounds such as music and laughter. Subtitles are timed text for people who can hear the audio but do not understand the language, so they carry the dialogue. The two file formats you will meet are SRT and WebVTT, and both are plain text that pairs each piece of text with a start time and an end time. A caption file starts as a transcript, and an automatic caption is a draft that needs checking before you publish it.

This guide explains the difference between the two, what the two file formats look like, what the accessibility rules require of captions, how to judge caption quality, and how to make captions from a transcript with free tools.

Captions serve viewers who cannot hear; subtitles serve viewers who do not understand the language

The web standard separates the two terms by the viewer's problem. Subtitles are a transcription or translation of the dialogue for when the sound is available but not understood ¹. Captions cover more: a transcription or translation of the dialogue, sound effects, relevant musical cues and other relevant audio information ². A subtitle file for a French film is for viewers who can hear the French and need English. A caption file is for a viewer who cannot hear the film at all.

The accessibility rules later in this guide are about captions. The short answer to the question of the difference between captions and subtitles is in its own page.

Closed captions can be switched off; open captions are burned into the picture

A closed caption is a separate text track. The viewer's player shows it only when the viewer activates it, usually with the remote or a menu option ³. The closed captions track can be switched off by the viewer.

An open caption is part of the video. Open, burned-in or hard-coded captions are visible to all viewers because they are embedded in the video ⁴. A viewer cannot turn them off. See open captions in the glossary.

You can burn a caption file into a video with FFmpeg. Its subtitles filter draws subtitles on top of the input video using the libass library ⁵. Burned-in captions are the safe choice when the player may not support a caption track. They are the wrong choice when viewers need to switch them off, or when you may need to correct a typo later, because the text is then part of the picture and the video must be re-rendered.

SRT and WebVTT are plain text files of numbered, timed lines

SRT is a simple caption format. Each entry is numbered, has a time range and holds the text, and the entries are separated by a blank line. The subtitles are numbered sequentially, starting at 1 ⁶. The timecode puts a comma before the milliseconds ⁷. A short SRT file looks like this:

1
00:00:01,000 --> 00:00:03,500
[Music playing]

2
00:00:03,600 --> 00:00:06,200
MARIA: The recording is ready.

Each numbered block is one caption. The first line is the counter, the second is the time range, and the rest is the text. The SRT file entry explains the format in more detail, and what an SRT file is and how it differs from VTT has its own short answer.

WebVTT is the version for the web. The W3C says that the SRT file format was used as the basis for the WebVTT text track file format ⁸.

Use SRT for uploads and desktop editing, and WebVTT where a browser plays the video. A converter can change one into the other, but you should check the result in the player you will use, because a separator or a line break that a strict parser does not accept will break the file.

WCAG requires captions for prerecorded video at Level A and for live video at Level AA

WCAG, the Web Content Accessibility Guidelines, sets two captioning criteria. Captions for prerecorded synchronized media are success criterion 1.2.2, and it is a Level A criterion ⁹. Captions for live audio are success criterion 1.2.4, and it is a Level AA criterion ¹⁰.

In the United States, the Justice Department's ADA Title II web rule ties state and local governments to these standards. Their web content and mobile apps must comply with Level A and Level AA success criteria in WCAG 2.1, unless compliance would be a fundamental alteration or an undue burden ¹¹. The compliance dates were moved by an interim final rule. Governments with 50,000 or more people must comply by April 26, 2027 ¹².

The FCC's rule for televised programming sets out caption quality in detail. It makes video programmers responsible for four standards, except for de minimis captioning errors ¹³. It is written for televised programming, so treat it as a benchmark for other video. The question of whether automatic captions meet the accessibility rules is answered separately, and this guide is not legal advice.

Good captions are accurate, synchronized, complete and placed where they do not hide content

The FCC names four quality standards: accuracy, synchronicity, completeness and placement ¹³. They make a usable checklist for any caption file, whatever tool made it.

Accuracy means the right words, in order, with the sounds included

The rule requires captioning to match the spoken words in their original language, in the order spoken ¹⁴.

Accuracy also covers what is not said. The FCC asks captioning to provide nonverbal information such as the identity of speakers, the existence of music, sound effects and audience reaction ¹⁵. A caption that only transcribes speech leaves out part of what the rule asks for, to the extent the program allows.

Synchronicity means the text appears when the speech does

Captions must begin when the speech or sounds begin and end about when they end ¹⁶. Delay matters most in live programs, where captions can lag the speech.

Completeness and placement

Captioning must run from the beginning of the program to the end, to the fullest extent possible ¹⁷. It must also not block important content on the screen, including character faces ¹⁸. A caption that covers a speaker's face, a name in the lower third or on-screen text can fail the placement standard even when the words are right.

Readability sets the timing and line length

Two practical limits are published by Netflix, a streaming service, in its subtitle guidelines. A subtitle event may last at most 7 seconds ¹⁹. For most Latin-alphabet languages a line holds 42 characters ²⁰. These are one company's standards, so use them as a benchmark for your own files, not as a universal rule.

Live captions are a different service

Live captioning is a separate service. A human version is called CART captioning, in which a transcriber types what is said as it is spoken. The WCAG criterion for live captions is Level AA, as noted above. Automatic live captions are a different trade-off, and the next section covers their accuracy.

Automatic captions are a draft: correct them before you publish

Automatic captions are fast, but they make word errors that you can measure. A common measure is word error rate (WER), the share of words that the recognizer gets wrong. In a 2025 test of one 7-minute-31-second episode, Whisper Large V3 Turbo had a word error rate of 1.5 percent ²¹. Apple's speech API had 10.3 percent ²², and NVIDIA's Parakeet v2 had 12.3 percent ²³.

Treat these numbers with care. They come from one episode, and the test's author says the models were evaluated the same way but that the exact numbers would differ under slightly different methods ²⁴.

WER alone does not tell you whether captions are usable. In a 2017 study that proposed a new caption metric, word error rate correlated only 0.11 with ratings of caption usability from deaf users ²⁵.

Viewer ratings can differ from error rates. In a 2026 online survey of 216 deaf and hard-of-hearing participants, live TV captions and automatic captions synchronized with the audio were rated similarly ²⁶. The participants rated clips from live TV, so the study does not show how automatic captions do in conversation. The practical lesson is to measure the captions as viewers will see them, not only by a word count.

Automatic systems also fail in specific ways. Whisper can produce text on sound that contains no speech. In a 2025 study, researchers played artificial non-speech sounds to the model, and a set of hallucinated phrases appeared frequently in those tests ²⁷. Platform-specific steps for adding automatic captions to short vertical videos are covered in the question on YouTube Shorts.

The working rule is simple. Use the automatic output as a draft, read it against the audio, fix names and terms, add speaker labels and sound tags, and check the timing before you publish.

When you review a draft, work through these checks in order:

1. Compare names, product names and technical terms with a written list.
2. Check numbers and dates against the audio.
3. Add a speaker name wherever the speaker changes.
4. Add sound tags for any sound the viewer needs, such as laughter or a phone ringing.
5. Check each cue's length and line length against the limits in this guide.

Playing the file at the speed viewers will see it is a good way to catch timing errors, because a caption that is a fraction of a second late is easy to miss on a read-through.

Captions start as a transcript and gain timing through alignment

A transcript has words but no times. A caption needs each piece of text placed on the timeline, and that step is called forced alignment: the known words are matched to the audio so that each gets a start and an end.

Whisper's own timestamps are a weak starting point for captions. They are per utterance rather than per word, and they can be inaccurate by several seconds ²⁸. WhisperX, a system built on Whisper, adds word-level timestamps with voice activity detection and forced phoneme alignment ²⁹. That is the method to look for when a tool promises word-level timing.

Making captions from a transcript with free tools

The free route is a local transcriber plus a text editor. The steps below use the Whisper command-line tool, which is free and runs on your computer.

1. Transcribe the audio. Run whisper interview.mp3 --model turbo --output_format srt in a terminal. The option that chooses the output format writes all available formats if you do not specify one ³⁰, so name srt or vtt to get one file. The transcription walkthrough for Whisper shows the full setup.
2. Check each cue against the limits. Keep each caption within the timing and line limits in this guide: no more than 7 seconds on screen ¹⁹, and 42 characters per line ²⁰ for Latin-alphabet text, if you follow Netflix's benchmark.
3. Correct the text while listening. Fix names, technical terms and numbers. Add speaker names where the speaker changes, and tags such as [Music playing] for sounds that matter to the scene. The FCC's standards ask for both ¹⁵.
4. Check the file format. Keep the SRT file for uploads and desktop tools, or convert it to WebVTT for a web player, and open it in the player you will use before you publish.
5. Burn in only when needed. If the captions must be part of the picture, FFmpeg can draw them onto the video ⁵.

If you do not want to install anything, free file transcription in the browser gives a draft without a terminal. The review steps still apply.

Where to go next

This guide covers caption formats, rules and quality. For the wider task of turning recordings into text, start with how to transcribe audio to text and video to text. Speaker labels and timestamps are covered in timestamps and speaker labels, and the choice between a machine draft and a human transcript is covered in human vs AI transcription. For the people captions serve, see speech to text for deaf and hard-of-hearing people in the accessibility guides. The full overview of the subject is in transcription.