Transcribe a meeting or lecture without uploading it
Drop in a recording and get a timestamped transcript you can read, search, or paste into an AI. It runs in your browser, so the file stays on your machine. No minutes cap and no account.
- No upload
- No minutes cap
- Audio or video
- Markdown, text or subtitles
Read straight off your disk. Works offline once the model is cached.
Why not just use Otter or Fireflies
For most meetings you probably should. Two things here are different:
The recording stays put
Any hosted service puts a copy of your recording on their infrastructure, under their retention policy. For a client call, a consultation or an HR interview that is often a risk you are not allowed to take. There is no server here to send it to. Open the network tab and check.
Nothing is metered
Free tiers elsewhere give you a few hundred minutes a month, often with a per-recording limit a real meeting blows through. Your hardware does the work here, so there is no meter to run. Do a three-hour workshop in two halves if you need to.
What they have that this does not: speaker labels, a bot that joins your calls, and a much bigger server-side model that handles jargon better. If you need those, go and use them.
Pasting a transcript into an AI
Most people transcribing a lecture are going to paste it into ChatGPT or Claude and ask for a summary, or to be quizzed on it. The Copy for an assistant button is for that.
It copies the transcript with a couple of lines in front saying the text came from speech recognition, and that the mistakes cluster in names, jargon, acronyms and numbers. This makes a real difference. Give a model a bare wall of text and it treats a misheard technical term as fact and builds on it. Tell it the source is imperfect and it flags the word instead. The note also explains the timestamps, so “when did they cover the deadline?” gets an answer you can go and check.
What it handles
- Audio or video — MP3, M4A, WAV, MP4, MOV, WebM and most else. A voice-memo recording of a lecture works the same as a screen recording of a call.
- Up to two hours per file, and no limit on file size. The two hours is a memory limit rather than a policy: the audio alone needs about 3.8MB per minute alongside the speech model. Longer recordings need splitting. All the limits, with the arithmetic.
- Sixteen languages, plus translation into English.
How long it takes, measured
This is the real trade, so here are numbers rather than adjectives. On a recent desktop with WebGPU — Chrome or Edge — transcription runs at about 8.6× faster than realtime:
- A 10-minute meeting: roughly 1 minute
- A 30-minute lecture: roughly 4 minutes
- A 60-minute lecture: roughly 8 minutes
Measured on a 31-minute recorded lecture, Chrome, 16GB Windows laptop, with whisper-base on WebGPU, on . Add about 15 seconds the first time for the model download, which your browser then caches.
On a machine without WebGPU it is much slower. The fallback runs on one CPU thread, and a long recording there is a wait measured in the recording’s own length or more. The tool tells you which path you are on before you start, rather than letting you find out forty minutes in. Progress is shown in minutes of audio processed, and you can stop partway and keep what it has done.
Accuracy is that of the smaller Whisper models: good on clear speech, weaker on heavy accents, crosstalk, and specialist vocabulary. There are no speaker labels. For study notes and searchable records it is more than enough; for a legal record it is a first draft.
Also need captions?
The same transcript can be burned into a video as captions, or exported as an .srt file. Switch to captions mode once it has run — it does not transcribe again.