Transcribe audio in your browser

Whisper tiny runs on your own CPU through WebAssembly. Drop a clip, get the text. The audio never leaves the tab, there is no server behind this page, and the model is cached after the first load.

no server, no upload
Nothing transcribed yet.

Ready. Choose a clip, drop one here, or use the sample.

load - transcribe - audio -

Models: Xenova/whisper-tiny.en at about 40 MB and Xenova/whisper-base.en at about 75 MB, the ONNX conversions of Whisper tiny and base, English only. Pick from the menu, tiny for speed and base when the audio is hard. Both run through transformers.js, cached after the first load, and only one is held in memory at a time. Measured on the bundled clip: on clean audio both models score the same, 2 word errors out of 17 from how 3:30 is written, on the noisy sample at 5 dB tiny turns moved into moves and base does not. Base costs about twice the load and twice the run. The sample is a sentence synthesized by tiny-voca, so a machine made it and a different machine reads it back. Expect small models to miss punctuation and proper nouns. Long clips are chunked internally and take proportionally longer.

Other demos on this account: background remover, browser transcription, tiny-voca text to speech, FLUX.2 klein inpainting.