Two things matter more than the software
Where the microphone sits, and whether people talk over each other. Both affect the result more than any setting on this page, and both affect every transcription engine equally — paid ones included.
A laptop microphone in the middle of a table picks up the two people nearest it well and everyone else as background. If a recording matters, put the recording device closer to whoever is doing most of the talking, or record each participant on their own device and transcribe the files separately. Crosstalk is the single largest cause of garbled output: when two voices overlap, the model does not produce two lines, it produces one confused one.
Room echo is the next worst offender. A hard-surfaced meeting room with a device in the centre gives the model a smeared signal to work from. Moving the device off the table onto something soft measurably helps.
Which quality setting to pick
Faster uses Whisper tiny, a roughly 39 MB download. More accurate uses Whisper base, roughly 73 MB. Both are cached by your browser after the first run, so the download is a one-off per setting.
Use the faster setting for clear speech in a quiet room, one or two speakers, and any recording where you mainly want to find the part you half-remember. Switch to the more accurate setting for anything with accents the model has to work at, technical vocabulary, several speakers, or audio recorded at a distance. The difference shows up most on proper nouns — names of people, products and companies are where the smaller model gives up first.
Neither setting is the largest Whisper model. If a recording is genuinely difficult — heavy background noise, a poor phone line — no browser-side model will rescue it, and the honest answer is to re-record rather than to keep re-running it.
How long it takes
The recording is processed in 30-second windows with a 5-second overlap between them, which is what makes long files work at all and stops words being cut in half at the seams. Where your browser supports WebGPU the model runs on your graphics card and is substantially faster; otherwise it falls back to WebAssembly on the CPU, which works everywhere and is slower. The page tells you which one it used while it runs.
Because it is your machine doing the work, the time scales with the length of the recording and with how busy your computer is. Closing other heavy tabs makes a real difference. A long recording will take longer here than it would on a server farm — that is the trade for the file never leaving your device.
Picking the language
Thirteen languages are listed plus automatic detection. If you know what language the recording is in, choose it rather than leaving detection on: detection is decided from the opening of the recording, so a meeting that starts with thirty seconds of English small talk before switching to Hindi will be detected as English and transcribed badly from there on.
Recordings that switch language mid-sentence — which is normal in a lot of workplaces — are the hardest case for any single-language model. Selecting the dominant language generally gives a better result than automatic detection does.
What this does not do
It does not label speakers. The output is continuous text, not "Speaker 1 / Speaker 2" — separating voices is a different model from the one that turns sound into words, and it is not running here. If you need attribution, the practical workaround is to have people say their name before they speak, or to record separately.
It does not add timestamps to the exported text, and nothing joins your call as a participant. For a meeting in progress there is a separate page — live meeting notes — which listens to the meeting tab and transcribes as it happens, still without uploading anything. It needs Chrome or Edge on a desktop, because tab audio is reachable nowhere else.
Getting the recording out of your meeting tool
An mp4 from a cloud recording works — the audio track is read out of it, so there is no need to convert anything first. mp3, m4a, wav, ogg and webm all work directly.
Step-by-step for each platform: Google Meet, Zoom, and Microsoft Teams. If you are recording a call you are on, it is worth knowing where consent is legally required before you press record — in some places every participant has to agree, not just you.
Is there a file size or length limit?
There is no limit imposed by us, because nothing is uploaded — the constraint is your own machine's memory. The recording is decoded into memory and processed in 30-second windows, so a multi-hour file will work but will take a while and will use a lot of RAM. If a very long recording fails, splitting it in half is the fix.
Why does the progress bar sit at zero when I start?
It only tracks the speech model download. Before that begins, the browser has to fetch the runtime that executes the model, and that stage has no progress reporting. It is downloading — it just cannot say how far along it is. This happens on the first run only; after that both are cached.
Does it work offline?
After the first run, yes for the model itself — your browser has cached it. The first run needs a connection to download the runtime and the model, roughly 65 MB on the faster setting or 100 MB on the more accurate one.
Can it tell speakers apart?
No. The output is continuous text with no speaker labels. Separating voices is a different model from the one that converts speech to text, and it is not running here. If attribution matters, ask people to say their name before speaking, or record each participant separately and transcribe the files one at a time.
Which file formats can I use?
mp3, m4a, wav, ogg and webm work directly, and so does an mp4 video — the audio track is read out of it, so there is no need to convert a screen recording first.
Why is my transcript full of mistakes?
Almost always the audio rather than the model. Overlapping speech is the biggest cause: when two people talk at once the model produces one confused line rather than two. Distance from the microphone and room echo are next. Switching to the more accurate setting helps most with proper nouns — names of people, products and companies — but no browser-side model rescues genuinely poor audio.
Should I use automatic language detection?
Only when you do not know the language. Detection is decided from the opening of the recording, so a call that begins with English small talk and then switches languages gets detected as English and transcribed badly from there. Choosing the dominant language explicitly gives a better result.
Can it transcribe a meeting that is happening right now?
Yes, but on a different page. This one works on a recording you already have. For a call in progress, live meeting notes transcribes it as it happens by listening to the meeting tab — no bot joins the call and the audio still never leaves your device. It needs Chrome or Edge on a desktop, because tab audio is only reachable there.