Tools › On-device AI
Auto subtitles and speech to text
Turn the speech in a video or recording into subtitles with OpenAI’s Whisper running in your browser. Fix any line, then download SRT, VTT or TXT, or burn them into the video.
The speech model downloads once to your device; your recording never leaves it. Nothing starts until you choose a file.
How it works
- The sound is decoded on this device and resampled to 16 kHz mono, the format Whisper listens to.
- Long recordings are cut into parts of up to 30 seconds at natural pauses, and each part is transcribed with timestamps. Lines appear as soon as each part is done.
- The spoken language is detected from the first speech, or pick it yourself. Translate to English turns speech in any of Whisper’s 99 languages into English subtitles.
- It runs on your graphics card with WebGPU when there is one, otherwise on your processor with WebAssembly.
Questions
Is my video uploaded?
No. The speech model downloads once from Hugging Face and is stored on your device; your file never leaves this tab.
How big is the download?
Fast is about 41 MB and Accurate about 77 MB without WebGPU. With WebGPU the graphics-card builds are about 120 MB and 206 MB. Both stay cached for next time.
How long does it take?
With WebGPU it is usually much faster than the recording plays. On a processor, Fast needs about half the length of the recording and Accurate about the full length, plus the one-time download.
Which model is used?
Whisper tiny (Fast) and Whisper base (Accurate) by OpenAI, released under the MIT licence, in the ONNX builds published by Hugging Face. You can use the subtitles commercially.
Can I put the subtitles on my video?
Yes. Burn into the video opens the subtitle burner with your lines and your video already loaded, ready to style and export.