What is offered
Quiet Transcript is provided free and as-is, with no warranty of accuracy, availability or fitness for a particular purpose. There is no account and no payment. This is a staging build: it may change or stop being available without notice, and nothing here is a service level.
How it is paid for, and what that does not buy
The site carries one advertisement, at the foot of the tool page, from Google AdSense. It is free to you and it is paid for that way. Nothing about the tool is behind it: there is no feature, no format, no length and no quality that depends on an advertisement being shown, and blocking it changes nothing you can do here.
Google is its own controller for what an advertisement request tells it about your visit, and what you agree with Google about advertising is between you and Google. Privacy and Cookies say what leaves your browser and what can be written into it. What does not travel is the material this site exists for: your recording, your transcript and your corrections are never sent to Google, to an advertiser or to anyone else, because they are never sent anywhere at all.
What using it puts on your device
The first time you transcribe something, the site sends your browser about 70 MB of speech model and WebAssembly runtime, and your browser keeps them so later runs do not download them again. The work then happens on your own hardware, at your own expense in time, battery and data. Clearing this site's data removes the download; Privacy says exactly where it sits.
Verify before you rely on it
Automatic transcription makes mistakes, and it makes them most where audio is worst — crosstalk, accents, distance from the microphone. A transcript produced here is a draft. Check it against the recording before quoting it, filing it or entering it into evidence. The tool plays the recording beside the transcript and lets you correct any line so that checking is possible rather than merely advised; the responsibility for doing it remains yours.
It is not a certified, sworn or verbatim transcript, and it is not a substitute for a qualified transcriber where one is required. Nothing here is legal, medical or professional advice of any kind.
Three limits worth knowing before you start
Ninety-nine languages, and weaker in most of them. The model is whisper-tiny, trained on ninety-nine languages and strongest in English. Outside English the draft is rougher, and in a language with little training data it may not be worth filing. The larger whisper-base model is offered on the tool page and is better, at an extra 80 MB to download.
The language is decided from the first thirty seconds. It is printed with the transcript and you can change it, which transcribes the recording again with your answer forced. A wrong guess does not arrive as an error — it arrives as a transcript — so check the language named beside the result before you rely on the words. The tool transcribes and never translates: what was spoken is what comes back.
No automatic speaker separation. Nothing here works out who is speaking, and this build will not gain it. You can type a label on any line, and those labels are written into whatever file you download — but they are yours, not a detection, and nothing verifies them.
Your responsibility
You are responsible for having the right to record and transcribe the material you process, and for meeting any consent or notification duty that applies where you are. The transcript is yours; so is what you do with it.
What it is built from
The speech model is OpenAI's whisper-tiny, licensed MIT, in the ONNX conversion published by onnx-community under Apache-2.0; whisper-base, offered as the larger option, is the same licence in the same conversion. They run through Hugging Face's transformers.js (Apache-2.0) on onnxruntime-web (MIT). Copies of those files are served from this site rather than from anyone else's network, and each one is pinned in this site's source by version and SHA-256.