No. Whisper runs inside your browser tab. The model weights are downloaded to you the first time, which is a download rather than an upload - your recording never leaves the device.

Because the transcription happens locally, the model has to be on your machine. It is cached afterwards, so the second transcription starts immediately. Tiny is a few tens of megabytes; large v3 turbo is over 1 GB, which is why it is not offered on phones.

Start with Tiny English if your audio is clear English speech - it is fast and usually enough. Move up to Base or Large v3 turbo for accents, crosstalk, background noise, or any language other than English, and choose the spoken language.

Whisper's multilingual models (Tiny multilingual, Base and Large v3 turbo) transcribe dozens of languages, and the page offers 30 common ones. You pick the spoken language in the Spoken language list, which starts at your browser's language; it does not detect the language by itself. Tiny English and the other English-only models understand English only. Base or Large v3 turbo are noticeably better than Tiny for languages other than English.

Yes. Whisper returns timestamps, so you can download the result as an SRT or WebVTT subtitle file as well as plain text.

Yes, if you ask it to. Tick Label speakers before you transcribe (or press Label speakers under a finished transcript) and a second pass works out who is speaking when. Each turn then starts with "Speaker 1:", "Speaker 2:" and so on in the text, the Markdown notes and the subtitle files (in WebVTT as a voice tag, <v Speaker 1>). You can rename a speaker, for example "Speaker 1" to "Dr Lee", and every download uses the name. It runs in your browser like the transcription, with two small open models (pyannote segmentation 3.0 and a WeSpeaker voice model, about 12 MB, downloaded once). Labels can be wrong: when people talk over each other only one of them is labelled, and two voices that sound alike can be merged into one speaker. If Auto finds the wrong number of people, set Number of speakers and the labels are regrouped straight away. Check the labels against the recording before relying on them.

It can lay the transcript out as notes. Download .md gives a Markdown file with a timestamped heading every 2 to 5 minutes (at a pause where there is one) and paragraphs split where the speaker paused. If you paste your slide titles, one per line, the headings use them instead: each title goes where its words are first heard after the previous one, and titles whose words are never heard are spaced evenly between their neighbours, marked as such on the page so you can check them. It does not summarise or rewrite anything.

Yes, with the Voice Recorder: record from your microphone, press Transcribe this recording, and this page opens with the recording loaded. The recording is passed through your browser's own storage, not uploaded.

There is no fixed limit. In current browsers the recording is read a piece at a time and fed to Whisper a minute or two at a time, so a two-hour voice memo needs about as much memory as a five-minute one. On a phone, recordings over 30 minutes are offered in 30-minute parts: each part's transcript can be downloaded when it is done, and the next part continues where it stopped. The time it will take is estimated before you start and measured on your device once it is running.

MP3, WAV, M4A, AAC, OGG, Opus and FLAC are read directly by the browser, as is the sound in an MP4, MOV or WebM video. For older or unusual formats the browser can't read, such as WMA, AMR, AIFF (in Chrome), AVI, WMV, FLV, 3GP, or AC-3 and ALAC audio, the page offers an extra converter: FFmpeg built for browsers, about 10 MB, downloaded once when you ask for it. It runs in your browser too, so the recording is still not uploaded. It holds the whole recording in memory, so it takes up to 4 hours on a computer and 90 minutes on a phone.

Yes, noise costs accuracy. Running it through the background noise remover first and then transcribing the result often works better than transcribing the raw file.