AI Video Editing Flows

Free auto captions: a local route and the free tools

You can caption a video for free on your own computer. We ran faster-whisper to get the words and FFmpeg to draw them, and measured both. Below are our word error rates and timings, then the free caption tools we reviewed, each with its licence and its catch.

The free local route we tested

Two free programs do the whole job. faster-whisper turns speech into timed words. FFmpeg then burns the captions into the picture. The audio stays on your machine. The first run downloads the speech model.

  1. Decode the audio with FFmpeg to 16 kHz mono samples. faster-whisper 1.2.1 could not open our files directly with the current PyAV, so we passed it samples.
  2. Transcribe with faster-whisper on the CPU: model small.en, int8, 8 threads, beam size 5, word timestamps on.
  3. Write one SRT cue per transcribed segment, and correct names and numbers by eye.
  4. Burn the SRT into the video with FFmpeg's subtitles filter, copying the audio as it is.

The full commands and the clips are on our faster-whisper test and our FFmpeg burn-in test.

What we measured

We ran two English models on the same three 60-second clips and compared each transcript with the film's official English subtitles.

ModelWord error rate, three clipsMedian time per clip (s)
small.en10.9% to 17.0%29.74 to 35.4
base.en19.0% to 26.6%7.83 to 40.51

small.en came closer to the subtitles on all three clips. The first small.en run took 175.43 seconds because it downloaded the model. Timings swing on a laptop that is busy with other work.

Part of each error rate is not a real mistake: the subtitles write numbers as words, spell a name differently and shorten some lines, so the figure likely overstates real errors. Treat it as a comparison between the two models, not an exact score.

Burn-in with FFmpeg worked. All 35 of 35 captions were in the picture at their time, in both the default style and the Shorts style, and the output kept the exact length of the input. Encoding took a median of 29.05 to 37.15 seconds per 60-second 720p clip.

Free caption tools with published cards

These are the free caption tools we have published cards for. We read each source and checked the licence on the card against the source. We did not run them, apart from faster-whisper, which we ran ourselves.

faster-whisper

Fast Whisper reimplementation that produces the transcripts captions are built from. It has an MIT file, and we tested it ourselves on three clips (see our faster-whisper test). The catch is GPU libraries on Windows can be fiddly. Licence on the card: MIT. Source. Checked 2026-10-17, reviewed from the source and not tested by us.

whisper.cpp

C and C++ port of Whisper that runs on a CPU, with demo media in the README and an MIT file. The catch is the models are large downloads. Licence on the card: MIT. Source. Checked 2026-10-17, reviewed from the source and not tested by us.

WhisperX

Whisper pipeline that adds word-level timestamps and speaker diarization, with a BSD-2 file. The catch is a GPU is recommended and diarization needs an access token. Licence on the card: BSD-2-Clause. Source. Checked 2026-10-17, reviewed from the source and not tested by us.

Buzz

Free desktop transcriber built on Whisper models. The catch is that its model downloads are large. Licence on the card: MIT. Source. Checked 2026-10-10, reviewed from the source and not tested by us.

Subtitle Edit

Offline open source subtitle editor for creating, syncing, converting and checking subtitles, with an MIT file and frequent releases. The catch is optional online translation and speech-to-text send data off the device. Licence on the card: MIT. Source. Checked 2026-11-06, reviewed from the source and not tested by us.

AutoSubtitles.com

Browser tool that transcribes a video, lets you edit text and timing, and burns styled captions into the export. The catch is free exports are watermarked and capped at 720p and 10 minutes, and SRT export is Pro only. Licence on the card: proprietary. Price on the card: Free $0 (10 min videos, 720p, small watermark, no signup); Pro price not read. Source. Checked 2026-10-10, reviewed from the source and not tested by us.

Remotion TikTok captions template

Official Remotion template that transcribes a video with whisper.cpp and renders TikTok-style word captions. The catch is there is no LICENSE file, Remotion terms apply, and the first run downloads a 1.5 GB model. Licence on the card: Remotion licence: free for individuals, up to 3 employees and non-profits; company licence above that. Source. Checked 2026-10-17, reviewed from the source and not tested by us.

Picking a route

  • You want an editable caption file: use a transcriber for the words, then a subtitle editor such as Subtitle Edit to correct and sync them.
  • You want word-by-word animated captions: you need word timestamps. WhisperX adds them, and the Remotion template renders TikTok-style words. Our own route is on animated captions with HyperFrames.
  • You do not want to install anything: AutoSubtitles.com runs in the browser, with the limits on its card.
  • You want the speed of a GPU: the cards for WhisperX and faster-whisper note GPU setup. We only ran the CPU.

For more caption tools and styles, see the captions workflow hub and caption styles.

Method and limits

  • Our figures come from three 60-second clips of acted film dialogue with music underneath. A clean talking-head recording would likely score better and a noisy one worse.
  • Error rates are measured against the film's official English subtitles after lower-casing and removing punctuation.
  • CPU only, on a laptop that was running other work. We did not test larger models, other languages or the voice-activity filter.
  • Every tool card other than faster-whisper is a source review, not a test. Check the source before you use a tool for client work.
  • Footage: Tears of Steel, (CC) Blender Foundation, mango.blender.org, used under CC-BY 3.0.

Page checked on Nov 4, 2026.