How to transcribe a recording without uploading it to anybody
An interview, a lecture, a meeting, a voice note, a consultation. Every transcription service wants the file, and the recordings people most need transcribed are the ones they should least be handing over. The model is small enough to run where the recording already is.
Written · Updated
Drop the audio or the video into SCRIBE, choose the spoken language or let it work that out, and transcribe. The model is Whisper (the small, fast size of it) running inside the tab through WebAssembly; the first run fetches the model once from this site and keeps it on your device, and after that the page works with the network off. The transcript comes out as plain text or as an .srt subtitle file with timestamps. The recording is never uploaded, and for a great many recordings that is not a preference but a requirement.
What is actually running
Whisper is a speech recognition model that was trained on a very large amount of multilingual audio, and it is unusually robust: accents, background noise, overlapping speech and bad microphones degrade it far more gently than the systems that came before. It comes in sizes, and the size decides both the accuracy and whether it can run on your laptop at all.
SCRIBE runs the smallest of them. That is a deliberate trade and it is the most important thing on this page: a small model is meaningfully less accurate than the large one a paid service runs on a server rack. Clear speech in a quiet room comes out very well. A four-person meeting on a laptop microphone comes out usable but wrong in places. Anyone telling you a browser tab matches a datacentre is selling something.
What it does well
- One person, close to the microphone, in a quiet room. Interviews, voice notes, dictation, a recorded lecture. This is the good case and it is most of what people transcribe.
- Thirteen languages, and it will detect which one rather than making you say.
- Timestamps. The
.srtit writes drops straight into a video editor or a player as subtitles. - Video as well as audio. The sound is taken out of the file for you.
- Long recordings, given time. It is not fast: a small model on a laptop is roughly real-time or a bit better, so an hour of audio is a coffee, not a second.
Where it goes wrong
- It does not label who is speaking. There is no diarisation: a two-person interview comes out as one column of text and you mark up the turns yourself. This is the gap people notice first.
- Names, jargon and figures. Proper nouns and technical terms are the first thing a small model loses, and it will confidently write something plausible instead. Read the numbers against the audio, always.
- Crosstalk. Two people at once produces one of them, or a blend.
- Heavy accents and poor recordings degrade it faster than they would the large model.
- It can invent. Whisper models are known to fill silence or noise with fluent, entirely fabricated sentences. This is documented behaviour, not a bug in this page. Long pauses and music are where it happens. Do not submit a transcript anywhere consequential without reading it against the recording.
Getting a better transcript
Most of the quality is decided before any software runs. A recorder near the speaker beats a better model further away; a hard-surfaced room is worse than a soft one; and the single biggest improvement available to most people is recording each speaker on a separate track, which also gives you the speaker labels the model will not. If a recording is long and only part of it matters, cut that part out with SNIP first. You will get the answer sooner and read it more carefully.
Why these recordings in particular should stay put
Think about what actually gets transcribed. A grievance hearing. A doctor's appointment. A therapy session. A research interview under an ethics approval that names who may hear the audio. A call with a client under legal privilege. A police interview. A journalist's source.
For every one of those, uploading the file is the disclosure: not a risk of one, the thing itself. The recording is on somebody else's machine, under their retention policy, subject to their jurisdiction and their subpoenas, and in more than one case used to train their next model. A university ethics committee will ask which processor you used and where the audio was stored. "It never left the machine" is the only answer that does not need a paragraph after it.
That is also the reason to be able to check rather than trust: load the page, disconnect the network, and transcribe. It still works.
Afterwards
A transcript is a document. If it is going anywhere, it may need the same care the recording did: ALIAS replaces names consistently before you paste anything into an AI, SENTINEL looks for card numbers and other things that should not be in it, and HUSH goes back to the recording itself and bleeps out the words you do not want in the copy you share. Subtitles that need retiming go to SUBS.
The short version
| The job | Here | Notes |
|---|---|---|
| Transcribe an interview or a voice note | Yes | The case it is good at |
| Keep the recording off every server | Yes | Model runs in the tab; works offline after the first run |
| Thirteen languages, detected automatically | Yes | |
| Subtitles with timestamps | Yes | .srt, for a player or an editor |
| Take the audio out of a video | Yes | |
| Label who is speaking | No | No diarisation; mark the turns yourself |
| Match a paid service's accuracy | No | A small model on your laptop, and said plainly |
| Be trusted without reading it | No | Whisper models can invent fluent text in silence |
| Transcribe an hour in a second | No | Roughly real time; it is your laptop doing it |
Questions people ask about transcribing a recording
How accurate is browser-based transcription?
For one person speaking clearly into a decent microphone, good enough to edit rather than retype. For a meeting on a laptop microphone with crosstalk, usable but wrong in places. It runs the small size of Whisper, which is meaningfully less accurate than the large model a paid service runs on a server. That trade is what buys you the file never leaving your machine.
Does it tell me who is speaking?
No. There is no speaker diarisation, so a two-person interview comes out as one column of text. If speaker labels matter, record each person on a separate track and transcribe them separately, which also gives a better transcript than any model would from a single mixed recording.
Can I trust the transcript?
Read it against the audio before it goes anywhere that matters. Whisper models are known to produce fluent invented sentences over silence, music or noise, and any small model will get names, jargon and figures wrong while sounding certain. Treat it as a first draft that saves you the typing, not as a record.
Is it safe to upload a recording to a transcription service?
It depends entirely on the recording, and the ones people most need transcribed (hearings, consultations, research interviews, privileged calls) are the ones where the upload is itself the disclosure. The file sits on their machine under their retention policy and their jurisdiction, and some services train on what they are given. If you would have to explain the choice to an ethics committee or a client, do not make it.
Does it work offline?
After the first run, completely. The model is fetched once from this site and kept on your device; from then on you can turn the network off, which is also the simplest way to satisfy yourself that nothing is being sent.
How long does an hour of audio take?
Roughly an hour, give or take, depending on the machine: it is your own processor doing the work rather than a rack of them. Cutting out the part you actually need first, with SNIP, is usually quicker than transcribing the lot.
Related tools
- Get the text out of a picture without uploading itWhat OCR reads well, what it reads badly, why handwriting does not work at all, and how to turn a scan into a PDF you can search.
- Merge, split and reorder PDFs without uploading themPut several PDFs into one, take one apart, pull pages out, turn and reorder them, number the pages, in the tab.
- Compress a PDF without uploading itA PDF is heavy because of its pictures, not its words. What downsampling, de-duplicating and tidying each save, which setting to pick, and the four kinds of file that will not shrink at all.
- Make a photo smaller without uploading itA phone photo is four megabytes and the form wants one. The two settings that make it small without anyone noticing, JPEG against WebP against PNG, the iPhone's HEIC, and what happens to the location.
- Make an EPUB from a Word file without uploading itAn EPUB is a zip of web pages, and a browser can write one: chapters at the headings, a cover, the details a reader shows, and what to check before sending it anywhere.
All 9 guides under “Doing the job without uploading it”. Part of the reading for clinics.
A letter, now and then
An occasional letter when there is something worth reading: the new guides and what changed in the tools. Never more than once a month and nothing else, and you can leave at any time on the unsubscribe page.