About TranscribeLocally
TranscribeLocally turns recordings into text without the recording ever leaving your computer. Not because uploading is always wrong, but because for a great many recordings it is the one thing you are not allowed to do.
Why it exists
Every mainstream transcription service works the same way: upload, wait, download. That is a sensible design and it produces good results. It is also, for a therapist, a solicitor, a journalist with a source, an HR investigator or a researcher holding a signed consent form, a disclosure of confidential material to a third party — one that has already happened by the time anyone reads a privacy policy.
The technology to avoid that arrived quietly. OpenAI published Whisper under a permissive licence, the community converted it to a format browsers can run, and WebAssembly and WebGPU made a browser tab fast enough to be practical. This site is that combination, packaged so it takes one drag-and-drop rather than a Python environment.
How it works, in one paragraph
The speech model — 207 MB of weights across three tiers — is served from this site and downloaded to your browser the first time you use it, then cached. Your file is decoded by the browser’s own media decoder, converted to the 16 kHz mono the model expects, and processed by your processor or graphics card inside a Web Worker. There is no upload endpoint on this site, because there is no server: it is static files on a CDN. Once the model is cached you can disconnect from the internet and it still works. The longer version is in how browser transcription works.
What it deliberately does not do
- No automatic speaker identification. Whisper has no diarization, and bolting one on would produce confident, wrong attributions exactly where transcripts matter most — during crosstalk. Speakers are labelled manually, in a couple of clicks.
- No fetching from YouTube or a URL. That needs a server to do the fetching, which would mean having a server, which would undo the entire premise.
- No huge models. The large Whisper checkpoints are more accurate and are multiple gigabytes. Nobody is downloading that to transcribe a voice note, so we ship the sizes that are reasonable and tell you plainly what they cost you in accuracy.
- No accounts, no storage, no history. There is nowhere to log in and nothing kept. Export what you want before closing the tab.
Who builds it
One independent developer, working on it outside a day job. That is worth knowing because it sets your expectations correctly: it is maintained attentively but not staffed around the clock, and there is no support desk. Bugs and corrections are genuinely welcome at hello@transcribelocally.com.
How it is paid for
Nothing today. It costs almost nothing to run because your computer does the work — the site serves static files and has no per-user compute cost at all, which is why there is no free tier to outgrow and no minutes to buy.
The intention is to carry advertising once there is enough traffic to be worth it. If that happens it will be display advertising in the page margins, and it will never involve reading, uploading or analysing your recordings — the architecture makes that impossible rather than merely forbidden. There are no tracking cookies today, which is why there is no cookie banner.
Being honest about the limits
A quantized base-sized model in a browser is less accurate on difficult audio than the large models behind paid services. For a noisy multi-speaker recording with no confidentiality constraint, a cloud service will do better and we would rather say so than pretend otherwise. The full account is in how accurate is it, really.
Nothing here is legal or compliance advice. The claim being made is narrow and technical: your recording is not transmitted anywhere. Whether that satisfies a particular obligation is a question for whoever advises you on it.