# Quiet Transcript

> Records and transcribes an interview, deposition or session inside the visitor's own browser.
> The recording is never uploaded, and there is no backend that could receive it.

The tool is complete and the site is published. It records from the microphone
in the page — the microphone is requested at the moment the visitor presses record and at no
other moment — and it also accepts a file they already have. Either way the page decodes the
recording, resamples it to the model's 16 kHz, and runs multilingual whisper-tiny in a Web
Worker, then offers the result as SRT, WebVTT or plain text. It also plays the recording in the
page — out of memory through Web Audio, with no media element and no URL — jumps to the moment
behind any line, and lets the visitor correct the words and type speaker labels that are written
into whatever file they download.

- Operator: AIErudit
- Contact: support@quiettranscript.com
- Home: https://quiettranscript.com/
- About: https://quiettranscript.com/about/
- Privacy: https://quiettranscript.com/privacy/
- Cookies: https://quiettranscript.com/cookies/
- Terms: https://quiettranscript.com/terms/
- Contact page: https://quiettranscript.com/contact/
- AI policy: https://quiettranscript.com/ai.txt
- Route inventory: https://quiettranscript.com/discovery.json

## What is true about this site

- No account, no upload, no server-side processing of visitor audio.
- This site sets no cookie of its own. Page views ARE counted, by a self-hosted cookieless Umami
  collector at analytics.pagefabrica.com that honours Do Not Track and never receives a file
  name, a duration or any part of a transcript.
- The site DOES carry advertising: Google AdSense, publisher pub-3015567014793107, declared in
  https://quiettranscript.com/ads.txt, in ONE box at the foot of the tool page below every
  control and every paragraph. Google's advertising library is not a script tag on any page — it
  is fetched only as that box nears the viewport, so a visit that does not scroll to the bottom
  makes no request to Google. When it is fetched, the request carries the visitor's IP address,
  user agent and page address, and Google may set or read advertising cookies; Google is its own
  controller for that. In the EEA, the UK and Switzerland Google shows its own consent message
  from a Google-certified consent platform, delivered through the same code; this site runs no
  consent surface of its own and does not store the answer.
- NOTHING about a recording ever reaches an advertiser: not a transcript, not a file name, not a
  duration, not the detected language. The audio and the transcript exist only in the page's
  memory and the advertising code is never given them. Blocking advertising changes nothing a
  visitor can do here.
- The speech model and its runtime are about 71 MB, served from this origin and kept by the
  visitor's own browser after the first run.
- Nothing is stored automatically. There is no autosave, no draft on a server and no setting that
  turns storing on; a reload loses the transcript and every correction.
- A RECORDING can be stored, and only by the visitor pressing "Keep on this device" for that
  recording. It is then written to this site's IndexedDB database in that browser, on that
  computer; it survives closing the tab and restarting the machine, it is never sent anywhere,
  and it is readable by anyone who can use that computer. The front page lists everything kept
  and deletes any of it, singly or all at once; clearing the site's data also removes it. A file
  the visitor loaded from their own disk is never offered for keeping.
- The default model is whisper-tiny, which understands 99 languages. The language is worked out
  from the first 30 seconds, PRINTED beside the transcript, and changeable by the visitor —
  changing it transcribes the recording again with that language forced. Detection from one
  30-second window is unreliable on a quiet or accented opening, which is why the answer is
  shown rather than assumed.
- The task is always transcribe and never translate. A recording in Russian comes back in
  Russian; this tool never turns another language into English.
- Accuracy outside English is materially weaker at this size. whisper-base is offered on the
  tool page as a deliberate upgrade, is better outside English, and downloads a further 80 MB
  from this origin only when a visitor chooses it. Nothing fetches it otherwise.
- There is no automatic speaker separation and this build will not gain one; any speaker label
  in a transcript was typed by the visitor.
- Transcripts are drafts and must be checked against the recording before being relied on. This
  is not a certified or verbatim transcript, and it is not legal or medical advice.