Skip to content
CaptionOnReelOpen editor

Whisper subtitle generator, in your browser

CaptionOnReel runs OpenAI's open-source Whisper speech model right in your browser, on your own GPU or CPU. You get word-by-word timings and animated captions burned into an MP4, without installing anything, without a server and without uploading your video.

Add captions to a video
  • Free to use
  • No signup
  • No watermark

Your video stays on your device: nothing is uploaded.

Whisper is OpenAI’s open-source speech recognition model, and it is very good at turning speech into text with timings. Normally you need Python, a command line or a paid service to use it. Here it runs as a web page.

What runs where

Everything below happens on your device; nothing about your video is sent to a server.

  • The models. Fast transcription uses Whisper base and High accuracy uses Whisper small, in versions that give a time for every word. They run through transformers.js, a library for running these models in the browser.
  • On the GPU. If your browser supports WebGPU, Whisper runs on your graphics chip (a 32-bit encoder and a 4-bit decoder, about 208 MB for base and 589 MB for small).
  • On the CPU. Without WebGPU, an 8-bit version runs on the processor instead (about 79 MB for base, 252 MB for small), split across up to three workers on computers and one on phones.
  • Downloaded once. The model is saved in your browser after the first download, so the next video starts right away.

What happens to your audio

  1. The sound is taken out of the video and converted to 16 kHz mono, the format Whisper expects.
  2. Long audio is split at natural pauses into pieces under 30 seconds. Whisper reads 30 seconds at a time, and cutting at pauses avoids the repeated or missing sentences you can get at fixed cut points.
  3. Each piece is transcribed with word-level timestamps, several pieces at once when there are spare workers.
  4. The word times are corrected: Whisper’s times run slightly late by a fairly constant amount per model (we measured about 160 to 170 ms), so that is taken off, and the first word after each pause is snapped to the moment the sound starts.
  5. The words become caption lines, ready to edit and style.

How accurate it is

We compared the output on a clear 61-second English clip with a reference transcript made by Whisper large-v3-turbo and word timings from a forced alignment:

Model and device Word error rate Words to fix
Whisper base (Fast transcription) 8.1% about 1 in 12
Whisper small, 8-bit, on the CPU 5.2% about 1 in 19
Whisper small, on the GPU (High accuracy) 4.0% about 1 in 25

After the timing correction, word starts were within about 50 ms of the real speech on average, and within about 120 ms for 9 words out of 10. On one test laptop, Whisper base on the GPU transcribed the 61-second clip in about 13 seconds.

That is one clip in one language, so treat it as a rough guide: noisy rooms, music, accents and mumbling all make any speech model less accurate. That is why every word can be corrected in the editor.

Languages

You can choose English, Spanish, German, French, Italian or Portuguese, or let it detect which of these is spoken. Whisper knows many more languages, but these six are the ones the small models write most accurately, so the others are left out rather than giving you poor captions.

Compared with running Whisper yourself

CaptionOnReel Whisper on your own computer
Setup Open a web page Install Python or an app such as Subtitle Edit
Models base and small Any size, including large-v3
Output MP4 with animated captions Subtitle files (SRT, VTT and others)
Styling 22 animated styles, fonts, colours Plain text; styling is up to your editor
Languages 6 All of Whisper’s languages

If you need subtitle files, very long recordings or the largest models, run Whisper on your computer. If you want a finished, styled video for Reels, Shorts or TikTok without installing anything, use the editor.

Questions

Is this the official OpenAI Whisper?

It runs OpenAI's open-source Whisper models (released under the MIT licence), converted to the ONNX format so a browser can run them. CaptionOnReel is not made by or affiliated with OpenAI.

Do I need a graphics card?

No. With WebGPU available, Whisper runs on your GPU; without it, or if the GPU fails, it runs on the CPU instead. Both give almost the same words.

Can I use Whisper large-v3?

No. The browser version uses Whisper base and Whisper small, which are small enough to download and run on phones and laptops.

Can I export an SRT file?

Not at the moment. The captions are burned into the MP4. If you need SRT files, a desktop app such as Subtitle Edit can run Whisper on your computer and save them.

Ready to add captions?

It takes a few minutes. Free to use, no signup, no watermark.

Choose a video