🔒Transcribe On-Device
Private speech-to-text with no server in the loop. Free, private, no signup — files never leave your browser.
First run downloads the AI model once, then it’s fast. 100% private — on-device. Fast = Whistle 16.9MB (7 languages) · Accurate = Whisper (99 languages).
Private by architecture.
The model runs inside your browser tab, so there is no server that could ever see the recording.
How to use Transcribe On-Device
Some recordings should never be uploaded to a third party: a source interview, a therapy session, a client call, a voice memo with a password in it. On-device transcription removes the question entirely — the speech model is downloaded once and then runs inside your browser tab, so the audio is decoded, transcribed and rendered without a request ever leaving the machine. This page is the same transcriber, framed around that privacy guarantee.
- 1Download the model once — 16.9 MB for Fast mode, larger for Accurate.
- 2Drop in the audio or video file you want transcribed.
- 3Choose a language, or let the model detect it.
- 4Copy the transcript or download it as .txt or .srt.
Looking for the full version of this tool? Open Audio to Text — it covers every option, not just this one.
Use cases
What transcribe on-device does
- ✓Speech model runs fully in the browser — no upload, ever
- ✓Word-level timestamps and language auto-detection
- ✓Works offline once the model is cached
- ✓No account, no quota, no watermark
Specs
- Input
- MP3, WAV, M4A, MP4 (audio track)
- Output
- .txt transcript and .srt subtitles
- Engines
- Whistle 16.9 MB or Whisper (99 languages)
- Privacy
- On-device inference, zero uploads
- Cost
- Free, unlimited, no signup
Frequently asked questions
What does on-device transcription actually mean?
The speech model is compiled to WebAssembly and executed by your browser, using your own CPU and GPU. The audio file is read from disk by the page, passed into the model in memory, and the transcript is drawn on screen. No API call is made, so there is no server that could log, retain or resell the recording — and the tool keeps working with no network at all after the model is cached.
Is it really private, or just private in theory?
The privacy policy on this site is short because there is little to describe: no accounts, no file uploads, no server-side processing. The one network request is the model download from a public CDN or model hub, and that request contains your IP address and the file URL, not your audio. Everything after the download happens in the tab.
How accurate is a 16.9 MB model compared with a large one?
The small model is tuned for clean, single-speaker speech in its seven supported languages and stays accurate for interviews and voice memos recorded in a quiet room. A large model wins on heavy accents, overlapping speakers, background music and rare languages. Because both run locally, you can pick per file without any privacy trade-off — the choice is about quality, not data.
Does it work offline?
After the model has been downloaded and cached, yes. The app shell itself is cached by a service worker, so the page loads without a connection and the transcription runs entirely on your device. That makes the tool usable on a plane, in a secure facility, or anywhere a network would be inappropriate.
More free tools
Need more control? Open the full video + audio editor →Was this tool helpful?