Audio Transcription

No — and the distinction is worth being precise about. Two things are downloaded on your first run: the Whisper model weights from the Hugging Face CDN, and the WebAssembly runtime that executes them from jsDelivr. Both are one-way downloads, cached by your browser afterwards. Your audio goes in the opposite direction only — it is decoded and recognised inside this tab, so the recording never leaves your device at any point.

Good, not perfect. On clear, close-mic English speech the small “Fast” model is genuinely useful. Accuracy drops on strong accents, background noise, crosstalk and overlapping speakers, and names, jargon and numbers are the usual mistakes. It also does not label who is speaking. Always read the transcript before relying on it.

The first run downloads the model — about 41 MB for “Fast”, 76 MB for “Higher accuracy” — and starts the WebAssembly engine. After that your browser has it cached, so later runs skip straight to transcribing. Transcription itself runs on your CPU, so a long recording takes real time and the tab needs to stay open.

The “Fast” model is whisper-tiny.en and handles English only — give it another language and you will get nonsense. The “Higher accuracy” model is whisper-base, which is multilingual, though quality varies a lot by language.