Transcribe audio to text

No. The file and the microphone recording are only read in your browser and recognised there with Whisper – they never leave your device, and neither does the text. The only downloads are the speech model (from Hugging Face) and the runtime (from jsDelivr), once; the browser keeps both for later visits.

“Fast” (Whisper tiny, about 42 MB) is enough for a quick overview and weak devices. “Standard” (Whisper base) is a good balance. “Accurate” uses Whisper large-v3-turbo (about 540 MB) on a graphics card and makes the fewest mistakes – in our test with a German recording about 2% wrong words versus about 6% with “Fast”. Without graphics-card support, “Accurate” is Whisper small and slower than real time.

The first time, the browser downloads the chosen model (about 42 to 540 MB depending on the level). After that it lives in your browser’s storage and starts in seconds. “Delete stored models” frees the space at any time.

With WebGPU (current Chrome and Edge, many Android devices) the graphics card does the work – a minute of audio takes a few seconds. Without WebGPU the processor does it, on a single core only: multiple cores would require security settings that do not work with this website. On a phone, a minute of audio may then take a minute or more.

Anything your browser can play: MP3, WAV, M4A/AAC, OGG/Opus, FLAC, WEBM and the audio track of MP4 videos. Whisper understands about 100 languages; the common ones can be selected, otherwise “Detect automatically” finds the language itself.

On a computer, recordings of an hour or more work; on a phone, rather up to about 30 minutes – the limit is the browser’s memory. For hours of material and the large models, a desktop version of Whisper is the better choice.

Very good for clear speech, but not perfect: names, technical terms, dialect, background noise or several people talking at once cause mistakes. Please proofread the text before sharing it.

Yes – switch on “Tell speakers apart” or click “Detect speakers” after the transcription. Two small extra models (pyannote segmentation 3.0 and WeSpeaker ResNet34, about 33 MB together, once) detect when someone speaks and group voices that sound alike. The tool works out the number of speakers itself; if you know it, set it – that is more reliable. You can give the speakers names, which then appear in every export too. If the detection is off, click a line’s time and pick the right speaker – or merge two speakers via the coloured dot. Passages where two people talk at once are marked. Hard cases are very short interjections, people talking over each other and voices that sound similar.

Yes. Choose “Microphone” at the top, click “Start dictation” and allow the browser to use the microphone. Speak in whole sentences: after every short pause what you said is recognised and appears as text – with a graphics card you see a preview while you are still speaking, without one the text follows a few seconds later. Whisper adds punctuation itself, but you can also say it (“period”, “comma”, “question mark”, “exclamation mark”, “colon”); say “new paragraph” to start a paragraph and “delete last sentence” to take the last sentence back. Which microphone is used you choose in the microphone box once there are several. While you speak, the “Standard” level at most types along, because “Accurate” needs several seconds per sentence – with “Accurate” selected, the recording is redone with it automatically after the dictation. Afterwards you can click any line and change it. The recording stays on your device: you can save it as WAV or have speakers detected. A dictation can be up to 60 minutes long on a computer and 30 on a phone.