AI Video Editing Flows

AI caption workflows for real footage

Captions are part of almost every video job. Compare built-in auto captions with code-based workflows and learn how to fix timing and spelling.

The first frame of the HyperFrames caption render over Tears of Steel
A frame from our own test run: We made animated captions with HyperFrames on a 60-second clip and timed every step.

Open the test run

The job and the deliverable

The job: put the spoken words on screen, timed to the speech. There are two common deliverables. A caption file (SRT or VTT) that a platform or player shows and lets viewers switch off. Or burned-in captions, drawn into the picture, often animated word by word.

Pick before you start. Burned-in captions look the same everywhere but cannot be turned off or fixed later without re-exporting. Caption files are easy to correct but each platform styles them its own way. Many teams export both.

What you need

  • Audio with clear speech. Captions inherit every problem in the audio.
  • A transcription engine that gives word timing if you want word-by-word animation.
  • A subtitle editor or text editor to correct the text.
  • The exact spelling of names, brands and technical words.
  • A style: font, size, colors, position, and the safe zone for the platform.

Recommended workflow

  1. Operator: decide caption file, burned-in, or both, and write the style rules (lines per caption, characters per line, position).
  2. Tool: transcribe. faster-whisper is a fast reimplementation of Whisper; whisper.cpp runs on CPU but its models are large downloads; WhisperX adds word-level timestamps, which animated captions need. The library notes faster-whisper's GPU libraries can be fiddly on Windows.
  3. Operator: correct the text. Subtitle Edit is an offline open-source editor for creating, syncing and checking subtitles. Its optional online services send data off your device, so leave them off for private footage.
  4. Tool: style and render. For animated captions, HyperFrames has caption blocks and an embedded-captions skill, and Remotion has an official TikTok captions template that uses whisper.cpp. We describe the HyperFrames route in Animated captions with HyperFrames, which is still in testing. Look at the finished styles in Animated caption styles.
  5. Operator: play the video with sound and watch the captions, not the picture. Check timing at the start, the middle and the end of the file.
  6. Tool: export the caption file and the burned-in version, then check sizes and the safe zone.

What we ran ourselves

We ran the free route on three 60-second clips of Tears of Steel: faster-whisper for the words, FFmpeg to burn them in, and a HyperFrames render for animated word captions. Each test has the commands, the timings, the memory and what failed: faster-whisper, FFmpeg burn-in, animated captions with HyperFrames and the Remotion TikTok template.

  • Word error against the film's official subtitles: 10.9% to 17% for small.en and 19% to 26.6% for base.en. Part of that is the subtitles' own wording, so use it to compare models, not as an exact score.
  • Burn-in with FFmpeg put every caption in the picture on all three clips, and a 60-second 720p encode took a median of 29.1 to 36.3 seconds.
  • The animated captions took a median of 3.7 minutes for one 60-second clip, most of it the HyperFrames render, with a peak of 1.7 GB of memory.
  • For burn-in and the animated captions we checked every caption in the picture by comparing frames with and without it. We did not test Hindi or other languages.

Not run by us: Submagic, AutoSubtitles.com's paid plan and any caption editor in the reviewed picks other than the four tools above.

Alternative route

Your editing app's own auto-captions are the simplest route when you only need a caption file and are happy to correct it by hand. We have not measured how they compare with code-based workflows; the benchmark page is a written protocol with no results: CapCut auto captions vs Premiere Pro vs a code workflow.

Hosted tools such as Submagic add animated captions to short clips in a web editor, and Descript captions from its transcript. Both are paid products; the library notes that Submagic's pricing cards and comparison table disagree on some limits. This route fits when you make many short clips and want a ready-made look.

Quality checklist

  • Every caption appears when its first word is spoken and leaves soon after its last word. Check three places in the video.
  • Spelling is correct for every name, brand and technical term. Read the captions aloud against the audio.
  • No more than two lines per caption, and line breaks fall at natural phrase boundaries.
  • Captions stay out of the platform's interface areas (safe zone) and contrast with the picture behind them.
  • Caption text is large enough to read on a phone at arm's length.
  • Burned-in version: exported at the final size and codec, with no captions cut off at the edges.
  • Caption file: opens in the platform's uploader without errors, and the language is set.

Known problems

Speech recognition makes errors on names, product terms and accents. We found no published accuracy numbers for Hindi or Indian-accented English, so assume those need a full read-through.

Transcription models can invent text over silence or music. Scan for caption lines that appear where nobody is talking.

Word timing drifts out of place if you edit the video after transcribing. Re-time after any cut.

Some dependencies in this area are archived or stale. The library research flags stable-ts (archived 2026-05-30), so check the repository status of any caption tool you adopt.

We ran faster-whisper, FFmpeg burn-in and a HyperFrames caption render on film clips, described under What we ran ourselves. The other caption tools in the reviewed picks were read, not run.

Tools and effects for this job

Pages on this site for captions: a workflow, a benchmark protocol and the caption style effects.

Reviewed picks from other creators

Transcription engines, subtitle editors and caption tools from other creators. Every card is a draft until the owner reviews it.

Source page screenshot, captured 2026-11-07

faster-whisper

By SYSTRAN on GitHub

Fast open source speech to text library that our own engines use for captions. The catch is that GPU libraries on Windows can be fiddly to set up.

Use it for: Generates the transcript that caption styles are built on.

Licence, price and checks
  • Licence MIT
  • Reuse keep the licence notice
  • Library check works, Oct 17, 2026

Source reviewed on Oct 17, 2026. Not tested by us.

Source page screenshot, captured 2026-11-07

whisper.cpp

By ggml-org on GitHub

C++ speech to text engine that runs locally and is widely used. The catch is that its models are large downloads.

Use it for: Transcribes audio locally for captions without a cloud service.

Licence, price and checks
  • Licence MIT
  • Reuse keep the licence notice
  • Library check works, Oct 17, 2026

Source reviewed on Oct 17, 2026. Not tested by us.

Source page screenshot, captured 2026-11-07

WhisperX

By m-bain on GitHub

Open source transcription tool that gives word level timings for captions and edits. The catch is that it wants a GPU, and speaker labels need an access token.

Use it for: Gets word timings that captions are timed to.

Licence, price and checks
  • Licence BSD-2-Clause
  • Reuse keep the licence notice
  • Library check works, Oct 17, 2026

Source reviewed on Oct 17, 2026. Not tested by us.

Source page screenshot, captured 2026-11-07

Subtitle Edit

By SubtitleEdit on GitHub

Mature free subtitle editor for fixing timing and text by hand. The catch is that its online translation and speech to text options send data off your machine.

Use it for: Corrects caption text and timing after automatic transcription.

Licence, price and checks
  • Licence MIT
  • Reuse keep the licence notice
  • Library check works, Oct 17, 2026

Source reviewed on Oct 17, 2026. Not tested by us.

Source page screenshot, captured 2026-11-07

Remotion TikTok captions template

By remotion-dev on GitHub

Official Remotion template for animated word by word captions. The catch is that it has no licence file, Remotion's company terms apply, and the first run downloads a large model.

Use it for: Builds animated short-form captions as code.

Licence, price and checks
  • Licence Remotion licence: free for individuals, up to 3 employees and non-profits; company licence above that
  • Reuse Remotion licence: free for individuals, for-profit organisations with up to 3 employees and non-profits; a company licence is needed for larger for-profits (LICENSE.md in remotion-dev/remotion, read 6 Nov 2026)
  • Library check works, Oct 17, 2026

Source reviewed on Oct 17, 2026. Not tested by us.

Source page screenshot, captured 2026-11-07

Buzz

By chidiwilliams on GitHub

Free desktop transcriber built on Whisper models. The catch is that its model downloads are large.

Use it for: Makes a transcript from a recording to caption from.

Licence, price and checks
  • Licence MIT
  • Reuse keep the licence notice
  • Library check works, Oct 10, 2026

Source reviewed on Oct 10, 2026. Not tested by us.

Designed card

Submagic (captions and short-form editing SaaS)

By Submagic

Paid captions and short-form editing service with credit based plans. The catch is that its pricing cards and comparison table disagree, and refund terms are missing.

Use it for: Adds styled captions to short vertical videos.

Licence, price and checks
  • Licence proprietary
  • Reuse editorial source
  • Price Starter USD $19/mo monthly or $12/mo yearly (45 credits, 15 videos, 2-minute videos); Pro $39/mo or $23/mo yearly (120 c
  • Library check works, Oct 10, 2026

Source reviewed on Oct 10, 2026. Not tested by us.

For teams that make this every week

If your team captions video every week, tell us about the job, the languages and where the time goes.

If your team makes this kind of video every week, tell us the job, the tools you use, how many you make a week and where it gets stuck: contact us. We read every message; there is no form to fill in and no account.

Page checked on Nov 4, 2026.