A Whisper model runs inside your browser and writes a recording out as text — nothing is uploaded
This browser cannot run the model
Recognising speech here needs WebAssembly, workers and the Web Audio API. Any current Chrome, Edge, Firefox or Safari will run this page. Everything else on the site works regardless.
The model
The model is downloaded once and kept by the browser, so a second visit starts straight away. The first download also brings the engine the models run on, about 21 MB, which both of them share. Everything is served from this site rather than fetched from anywhere else, and once it is here the recognition itself needs no network at all.
The recording
Drop a recording hereor click to choose one — MP3, WAV, M4A, AAC, OGG, FLAC, or the soundtrack of a video, up to 10 minutes
The file is read straight from the disk and never leaves the browser. Anything the browser can play will do, including the soundtrack of a video; it is brought down to the single sixteen-thousand-a-second channel the model was taught on before it is read.
Settings
Naming the language is worth doing whenever you know it: left to work it out, the model listens to the opening half-minute and decides, and a quiet or accented beginning can send it down the wrong language for the whole recording. Asking for the times also gives you a subtitle file.
Recognition
The work happens on a thread of its own, so the page stays usable while it runs and stopping takes effect at once. Recognition is arithmetic rather than a lookup: how long it takes depends on the machine, and a phone will be several times slower than a laptop.
The text
The text is yours to edit before you save it — recognition gets punctuation and proper names wrong often enough that a read-through is always worth the minute. Subtitles are offered only when the times were asked for.
About turning speech into text
Writing down what someone said used to be a trade. A shorthand typist could keep pace with a speaker at something like two hundred words a minute, and until the end of the twentieth century that was the only reliable way to get a meeting, a hearing or an interview onto paper. The first machines that tried it wanted a pause between every word and a training session with each new voice, and they were still wrong often enough that checking the result took about as long as typing it from scratch.
What changed is that recognition stopped being a matter of matching sounds against a dictionary and became a matter of prediction: a model trained on hundreds of thousands of hours of ordinary recorded speech, in dozens of languages at once, guessing the next word from everything it has heard so far. That is why recognition now punctuates, spells names it has read before, copes with accents and background noise, and can even listen in one language and write in another. It is also why it is confidently wrong sometimes — it is guessing what fits, and a plausible wrong word can beat an unlikely right one.
Most of the practical difference comes from the recording rather than from the model. A microphone close to the speaker, one person talking at a time, and a room without a television on will each do more for accuracy than any setting on this page. Where the text is going matters too: notes to yourself can be left as they come out, but anything meant to be read by someone else, quoted, or kept as a record deserves a pass by eye — the mistakes recognition makes are rarely nonsense, which is exactly what makes them easy to miss.