Turn a short English recording into editable captions, right here on your device.
Whisper tiny.en · English only · q8 CPU/WASM · size loading model + runtime assets.
Whisper MIT; ONNX conversion/runtime notices included. License · How local inference works
Tested on desktop Chrome with 16 GB RAM; observed browser working set about 1.4 GB. Low-memory/mobile devices and WebGPU are unvalidated.
WAV, MP3, M4A, OGG, WebM audio, FLAC · browser dependent · 50 MiB / five minutes max
Choose audio, load the model, then transcribe.
You stay in control of the final captions.
Text edits and valid timing adjustments are used in the export. Times must be ordered, nonoverlapping, and inside the recording.
One English model, one inference worker. Audio is decoded locally, downmixed/resampled to 16 kHz, then processed in overlapping 30-second windows. Silence below an RMS threshold of 0.0001 is skipped; quiet speech can be missed, and hallucinations can still occur.
Load model fetches assets from this host. Cache assets explicitly stores only reusable model/runtime files for offline inference in an already loaded page; reopening the site offline is not guaranteed. Clear session unloads the model and clears audio/caption references. Remove model cache deletes only this app’s opt-in asset cache.
Five-minute limit is provisional. CPU speed and RAM vary; keep recordings short on constrained devices. Unsupported decoding, WASM, or model errors never send audio to a cloud service.
WebVTT export starts with the WEBVTT header and uses millisecond dot timestamps. Caption text markup characters are escaped. Review the downloaded result in your player before publishing; model segment boundaries are approximate.