Auto-Captions and Kinetic Captions

Auto-captions are captions generated automatically from a video's speech, timed so that each word or phrase appears as it is spoken. Kinetic captions are the animated style that dominates short-form video: words that pop in one at a time, highlight as they are said, or grow for emphasis. Both are built on speech recognition with word-level timestamps. They serve two audiences at once. Captions are an accessibility requirement for deaf and hard-of-hearing viewers, and they are also a retention tool for the large share of people who watch social video with the sound off. They have become one of the most widely used features of AI video editing.

How Auto-Captions Work

Transcribe the audio with word timings and, ideally, speaker labels from speaker diarization.
Segment the words into caption units. Short phrases suit kinetic styles, and one or two lines of readable length suit subtitles. Good segmenters break at natural pauses and grammatical boundaries rather than mid-phrase.
Time each unit to the speech, with minimum durations so short phrases stay on screen long enough to read.
Style the text: font, size, color, background box, position, animation and emphasis.
Deliver the captions either burned into the picture or as a separate caption file.

Every step depends on the timestamps. A recognizer that places words half a second early makes captions that run ahead of the speaker. One that invents text over silence or music, a known failure of some speech models, puts words on screen that nobody said.

Caption Styles

Subtitles. One or two lines at the bottom of the frame, the convention of film and television.
Kinetic, word by word. Each word appears or highlights as it is spoken, often centered in a vertical frame. This is the default look of creator video on TikTok, Reels and Shorts.
Emphasis. Key words are enlarged, colored or animated to carry the point of a sentence.
Boxed. Text on a solid or translucent box, which improves legibility over busy footage.
Minimal. Small, clean type for brand and documentary work where captions should not dominate.

Style is a creative choice, but legibility is not optional. High contrast, sizes readable on a phone, and on-screen time long enough to read matter more than animation.

Burned-In Versus Sidecar Captions

Burned-in (open) captions are rendered into the video pixels. They always show, look identical on every platform and can be styled freely, which is why kinetic captions are almost always burned in. But viewers cannot turn them off, resize them or switch language, and the text is not machine-readable.

Sidecar (closed) captions travel as a separate file that the player displays. Viewers can toggle them, and platforms and assistive technology can restyle them. Several languages can ship side by side for localization, and search engines and AI systems can read the text, which matters for video GEO. Two formats dominate the web:

SRT, from the SubRip program. It is a simple numbered list of time ranges and text with no formal standard, and it is accepted almost everywhere.
WebVTT, a W3C specification from the Timed Text Working Group (a Candidate Recommendation Draft, last updated May 2026). It is the native caption format of HTML5 video and supports positioning and styling.

Broadcast and streaming delivery often require other formats, such as TTML-based profiles, but SRT and WebVTT cover most online publishing. Many creators ship both kinds: burned-in kinetic captions for the feed, plus an uploaded SRT for accessibility and search.

Safe Areas and Layout

Vertical platforms overlay their own interface on the video: captions, usernames, buttons and progress bars, mostly along the bottom and right edges. Captions placed there get covered. The practical rule is to keep captions inside the platform's safe area, which on vertical video usually means the middle band of the frame. Captions must also avoid faces, which computer vision can locate, on-screen text and lower-third graphics. Captions that move between shots to stay clear of the speaker are a sign of a careful editor or a good tool. When footage is reframed from landscape to vertical (see auto reframe), caption placement has to be rechecked too.

Accessibility

Under the W3C's Web Content Accessibility Guidelines, captions for prerecorded video are a Level A requirement (Success Criterion 1.2.2), the most basic level of conformance. Accessible captions do more than transcribe dialogue:

They identify speakers when it is not obvious who is talking.
They describe meaningful non-speech audio, such as [music], [laughter] or [door slams].
They stay on screen long enough to read.

Auto-generated captions are a starting point. YouTube, which generates automatic captions for uploads, itself cautions that their accuracy varies. A human pass over names, technical terms and anything legally or medically significant is still needed for captions to be both accurate and accessible.

Tools in 2026

Captioning is built into nearly everything:

Platforms. YouTube, TikTok (auto captions since April 2021) and other social platforms caption uploads themselves.
Creator apps. CapCut and Captions, from the company now called Mirage, made kinetic styles mainstream.
Professional editors. Final Cut Pro 12.3 (June 2026) added Generate Captions on Apple silicon, initially for US English. Premiere, DaVinci Resolve and Descript all generate captions from their transcripts.
Agentic and code-based tools. These generate captions from the same transcript the edit is cut from, so captions stay in sync when the cut changes. They also write an SRT next to every export (see video as code and agentic video editing).

Getting Captions Right

Correct the transcript before styling, especially names and brand terms.
Check captions over music and silence for words that were never said.
Break lines at natural phrase boundaries, and keep one idea per caption.
Preview on a phone, inside the platform's interface, before publishing.
Upload a sidecar file even when captions are burned in.

For the transcript workflow captions come from, see transcript-based editing.