Whisper AI · runs on your device · nothing uploaded
Audio to text that never leaves your laptop
Drop in a recording and get a clean, timestamped transcript, subtitles or meeting notes. The speech model runs inside this browser tab — no upload, no account, no per-minute pricing, and no third party with a copy of what was said.
- 207 MB of models, self-hosted
- Works offline after the first run
- No file-size limit but your own RAM
Language and translation need the Multilingual model.
Drop an audio or video file
or , paste with Ctrl + V, or
- MP3
- MP4
- WAV
- M4A
- MOV
- WEBM
- OGG
- FLAC
- AAC
The file is read by this page and never uploaded — no account, no size limit but your own RAM.
Download
Click a line to jump the audio there. Click the text to fix a word — your edits go into every download.
How it works
- Drop an audio or video file onto the page, or record straight from your microphone.
- The speech model downloads once (81 MB for the default) and is cached in your browser.
- Transcription runs on your own processor or GPU — the file never leaves the page.
- Fix any word by clicking it, then download plain text, SRT, VTT, Markdown, CSV or JSON.
What you get out
-
Readable text, not a wall
Segments are grouped into paragraphs at natural pauses, with optional timecodes. Copy it straight into a document.
-
Subtitles that pass a player
SRT and VTT with cues split at sentence boundaries and wrapped to two readable lines — not one 40-second block of text.
-
An editor, not a text dump
Click a line to jump the audio there; click the words to fix a name the model misheard. Every export uses your corrections.
-
99 languages, or English out
Switch to the multilingual model to transcribe in the language spoken, or translate the speech directly into English text.
Why do it in the browser at all?
Every mainstream transcription service works the same way: you upload the recording, it is processed on someone else’s servers, and a copy sits in an account you have to remember to delete. For a podcast that is fine. For a therapy session, a witness interview, a client call or a discussion about someone’s medical history, it is a disclosure — and often one you are not allowed to make.
This tool exists for that second case. The speech recognition model is downloaded to your machine and run by your own processor. There is no upload endpoint on this site to send a file to, which is a stronger promise than a policy: you can open your browser’s network tab, watch the whole transcription happen, and see that nothing goes out. Once the model is cached you can switch off Wi-Fi entirely and it still works.
The honest trade-off: your laptop is slower than a datacentre, and a quantized model running in a browser is a little less accurate than the paid cloud APIs on hard audio. If your recording is already public, a cloud service will be faster. If it isn’t, this is the version that doesn’t make you choose between convenience and confidentiality.
Which model should you pick?
| Tier | Download | Languages | Best for |
|---|---|---|---|
| Fast Smallest download, quickest run | 45.1 MB | English only | Voice memos, clear single-speaker dictation, a first pass to see whether a recording is worth transcribing properly. |
| Balanced The default — better words, still small | 81.1 MB | English only | Interviews, meetings, lectures, podcasts — anything you intend to read afterwards. |
| Multilingual 99 languages, or translate into English | 81.3 MB | 99 + translate | Non-English recordings, bilingual interviews, or getting the gist of a foreign-language clip in English. |
Sizes are the exact bytes your browser downloads, not a marketing figure. Bigger checkpoints (small, medium, large) exist and are more accurate — we don’t ship them because a quarter-gigabyte download to transcribe a ten-minute clip is a worse deal than it sounds, and even then it isn’t court-grade. For anything that has to be exact, budget for a human.
Frequently asked questions
Is my recording really not uploaded?
Yes, and you can check rather than take our word for it. Open your browser’s developer tools, switch to the Network tab and transcribe a file: you will see the page and the model weights come down, and nothing go up. There is no upload endpoint on this site to send a file to. Once the model has been cached you can disconnect from the internet entirely and transcribe with the network off.
Why is there a download the first time?
Because the speech recognition model has to be on your machine to run on your machine. The default English model is about 81 MB and the smallest is 45 MB. It is downloaded once, stored in your browser’s cache, and reused for every file afterwards — including offline. Clearing your browser storage removes it, and it will download again next time.
How accurate is it compared with Otter, Rev or Descript?
On clean, close-microphone English speech it is close enough that you will mostly be fixing proper nouns. On hard audio — heavy accents, several people talking over each other, a phone on a table in a big room — the paid cloud services are noticeably better, because they run much larger models than a browser can reasonably download. Expect to proofread. No automatic transcription of any kind, cloud or local, is accurate enough to file without reading it.
Can it tell who is speaking?
Not automatically. Speaker diarization is a separate model that Whisper does not include, and pretending otherwise would produce confidently wrong labels. Instead you can label speakers manually: click a line, name the speaker, and the label applies to that line and every one after it until the next label. For a two-person interview that is usually a handful of clicks, and the labels appear in every export.
What file formats work?
Anything your browser can play — MP3, WAV, M4A, AAC, FLAC, OGG, Opus, MP4, MOV, WebM and more, audio or video. The browser’s own decoders do the work, so the exact list depends on your browser and operating system. A few container/codec combinations (some MKV, WMA and AMR files) have no browser decoder; converting to WAV or MP3 first solves it.
How long can the recording be?
There is no imposed limit — the constraint is your device’s memory, because the whole decoded audio has to be held at once. Files up to about an hour are comfortable on a typical laptop; multi-hour recordings work but use a lot of RAM and take a while. If a long file makes the tab unstable, split it and transcribe the pieces.
Does it cost anything, and what is the catch?
It is free, with no account and no per-minute charge, because it costs us nothing to run — your computer does the work, not our server. The site is intended to carry advertising later. It does not run ads or any tracking cookies today.
Which browser and hardware do I need?
Any current Chrome, Edge, Firefox or Safari on a desktop or laptop. Browsers with WebGPU (recent Chrome and Edge) run several times faster because the model uses your graphics card; everything else falls back to the processor, which works but takes longer. Phones can run the Fast model on short clips, but a laptop is a far better experience.