Transcribe sensitive recordings — therapy notes, legal calls, medical dictation — with AI that runs entirely on your device. No server ever hears a word.
Files are processed locally and never uploaded.
Open this page and let the Whisper model load once — after that first download, everything runs on your device.
Drop your recording into the uploader above (MP3, WAV, or M4A). It is decoded and transcribed locally.
Copy the transcript or download the TXT, then store it with the same care you'd give any sensitive document.
A voice recording carries two layers of sensitivity at once: the content of what's said, and the biometric identity of who's saying it. Therapy sessions, legal client calls, medical dictation, HR conversations — the recordings people most need transcribed are exactly the ones that should never sit on a third party's servers, subject to someone else's retention policy, breach risk, and business model.
The standard transcription workflow asks you to accept that trade silently: upload first, ask questions never. This page exists for everyone who'd rather not make it. The audio never leaves your machine because there's simply nowhere for it to go — no account, no upload endpoint, no processing queue.
The engine is OpenAI's Whisper — the same model family behind many commercial transcription products — running directly in your browser via WebAssembly (ONNX runtime). On your first visit, the model files download once and are cached in your browser; from then on, your audio is decoded and transcribed by your own CPU, and the text appears as if by magic.
Skeptical? You should be — privacy claims are cheap. Verify it: open your browser's network inspector, drop in a recording after the model has loaded, and watch the network tab stay silent. After that first model download, you can even disconnect from the internet entirely and transcription keeps working. That's not a metaphor for privacy; it's the mechanism.
Let's be precise, because precision is the point. Covered: your audio file is never transmitted anywhere; there's no account tying transcripts to you; no server logs your usage. The only network activity is the one-time download of the model weights themselves — standard files from a CDN, containing no information about you or your recordings.
Not covered: the security of your own device (encrypt your drive, lock your screen), and the afterlife of the transcript — once you download the TXT or copy the text, protecting it is your workflow's job, same as any sensitive document. Local transcription removes the cloud-transfer risk completely; it doesn't replace basic device hygiene.
Therapists turning session recordings into clinical notes. Lawyers transcribing client calls without routing privileged conversations through a vendor. Doctors dictating notes that contain patient information. HR professionals documenting sensitive conversations. Journalists protecting sources. Anyone whose recordings carry legal, ethical, or personal weight beyond the ordinary.
For these users, the accuracy story is secondary — Whisper's three tiers (tiny for speed, base for balance, small for precision) handle the transcription quality — and the privacy architecture is the product. When the question is "can I use a transcription tool for this at all," on-device is the answer that makes it a yes.
Removing the cloud transfer eliminates the biggest single risk in most privacy frameworks — but no tool can certify your whole workflow compliant. Whether you're thinking about HIPAA, GDPR, attorney-client privilege, or your institution's ethics board, local processing is a strong foundation, not a complete answer: your device security, storage practices, and consent procedures still matter.
What we can say plainly: with on-device transcription, there's no vendor to vet, no data processing agreement to sign, no subprocessor list to audit — because there's no vendor in the loop at all. For many professional workflows, that simplicity is worth more than any compliance badge.
No — it's a standard download of AI model weight files from a CDN, identical for every user. It contains no information about you, and your audio is never part of it. After that first download, transcription works fully offline.
Yes, after the first visit. Once the Whisper model files are cached in your browser, disconnect from the internet and everything keeps working — upload a file, transcribe, export the TXT. Nothing needs the network again.
Local processing removes the cloud-transfer risk, which is the hardest part of most compliance puzzles — there's no vendor, no data processing agreement, no subprocessor to audit. But compliance covers your whole workflow (device security, storage, consent), so treat this as a strong foundation, not a legal guarantee.
It's gone unless you saved it — transcripts live only in the page's memory. Copy the text or download the TXT file before closing, and store it with the same care you'd give any sensitive document.