Transcription data flow
Updated 2026-10-25 · for ethics protocols, DPIAs and IT reviews
This page describes what happens to a recording in the interview transcriber. It is written to be quoted in a research protocol, a data protection impact assessment or an IT review. It covers the transcriber only: AudioRinse's Claude connector works differently and is described in the privacy policy. The site's other transcription and subtitle pages (transcribe audio, add subtitles to a video and the rest) are the same transcriber with the same data flow; a video with subtitles burned in is also made in the browser.
Statement
AudioRinse's interview transcriber (audiorinse.com/interview-transcription) processes recordings entirely on the user's own computer. The audio file is decoded and transcribed by an open-source speech recognition model (OpenAI Whisper), and speakers are identified by open-source models (pyannote segmentation 3.0 and NVIDIA TitaNet), all running inside the web browser. The recording, the transcript, the speaker names and the file name are never transmitted to AudioRinse or to any third party. Transcripts are stored only in the browser's local storage on that computer and can be deleted there. The models are downloaded once from Hugging Face; after that the transcriber works without a network connection.
What never leaves the computer
- The audio or video file, and the audio decoded from it.
- The transcript, including any corrections and speaker names.
- The file name.
Every network request the transcriber makes
| To | What and when | Data it receives |
|---|---|---|
| audiorinse.com (hosted by Vercel) | The page, its program code and the transcription engine (onnxruntime-web, sherpa-onnx), once; then from the browser's cache. | Standard request logs: IP address, browser, the page requested. |
| huggingface.co (Hugging Face's CDN) | The model files, the first time only: Whisper small (249 MB) or Whisper large-v3-turbo (about 560 MB), and the speaker models (46 MB). Public files at fixed versions. | IP address and which public model file was downloaded. No audio and no transcript. |
| audiorinse.com/api/export | Whether the visitor has a Pro plan, when the page opens and before a Pro download. | The account (if signed in) or a device cookie. No audio and no transcript. |
| Clerk and Stripe | Only when the user signs in or buys Pro. | Email address and payment details, as for any account. Never audio or transcripts. |
| Usage statistics (Vercel Web Analytics; PostHog where enabled) | Page views, and events such as "transcription finished". | The page, browser type, approximate country; for events, the length in broad buckets (e.g. 30 min+) and the speed. Never file names, audio or text. |
Nothing about the recording is sent while it is transcribed: the only requests then are the usage events above, which carry no audio and no text. You can watch this in the browser's developer tools (Network tab).
Where transcripts are stored
In the browser's own storage (IndexedDB, database audiorinse-transcripts) on that computer only. They are listed under the transcriber as “Your transcripts, on this device” and can be deleted there, or all at once by clearing the site's data in the browser settings. While a long recording is being transcribed, the progress so far is saved there too, so an interrupted transcript can continue; it is deleted when the transcript is finished. The recording itself is not stored.
Checking it yourself: the Wi-Fi-off test
- Open the transcriber and transcribe a short recording once. This downloads the model.
- Turn off Wi-Fi, or unplug the network cable.
- Reload the page and transcribe the interview. It works the same, because nothing needs the network.
The models
- Speech recognition: OpenAI Whisper large-v3-turbo (MIT), ONNX build
onnx-community/whisper-large-v3-turboat revision360ebcd; or Whisper small (Apache-2.0),onnx-community/whisper-smallat36050c4. - Speaker detection: pyannote segmentation 3.0 (MIT) and NVIDIA NeMo TitaNet-small (NeMo Toolkit licence, Apache-2.0), run by sherpa-onnx (Apache-2.0).
- None of them was trained on your data, and none is trained by using the transcriber.
Questions
For data protection questions or a signed copy of this statement, write to privacy@wikivec.com. AudioRinse is operated by Wikivec Inc., 1111B S Governors Ave STE 7867, Dover, DE 19904, United States.