Transcribing recordings you are not allowed to upload
Updated
In-browser transcription runs the speech model on your own machine, so the recording never leaves your device and there is no third party to trust with it. You can verify this directly: open your browser's network tab while you transcribe and watch what crosses the network — the model comes down, nothing goes up.
The problem with every hosted transcription service
The material that most needs transcribing is often the material you are least permitted to send anywhere: a recorded deposition, a clinical consultation, an HR interview, an internal board discussion. Uploading it to a web service creates a copy on someone else's infrastructure, governed by their retention policy, in a jurisdiction you may not have chosen.
The usual answers are unsatisfying. An enterprise agreement moves the risk rather than removing it. Running a model locally is possible but demands the skills and hardware to install and operate one.
What in-browser transcription actually changes
Your browser already contains everything needed: a video decoder to get at the audio, and — on recent hardware — a GPU compute interface that can run a speech recognition model. Given both, transcription can happen inside the page, with the file read directly from your disk.
The distinction that matters is the direction of travel. The model weights are downloaded to you, once, and cached. Your recording moves in the opposite direction from a hosted service: it does not move at all.
How to verify it rather than believe it
A privacy claim you cannot check is a marketing claim. This one is checkable, and it is worth doing once before you trust a tool with anything sensitive.
- Open your browser's developer tools and select the Network tab
- Load the tool, drop in a file, and transcribe it
- Sort by size: the only large transfers should be the model files, arriving from a Hugging Face domain
- Look for any request with a large upload payload. There should be none
- For an even stronger test, disconnect from the network after the model has downloaded. Transcription still works
The practical limits, stated plainly
Transcription runs at your machine's speed. With a reasonably recent desktop GPU that is a fraction of the video's length; on a CPU it is roughly its length or more. For a long recording that is a real wait, and it is the honest price of the file not going anywhere.
Accuracy is good but not court-ready. The models available in a browser are the smaller whisper builds, and they make mistakes on proper nouns, technical vocabulary, overlapping speakers and heavy accents. Anything consequential needs a human pass — which is why the editor flags the lines most likely to be wrong rather than presenting the transcript as finished.
There is no speaker identification. If who said what matters, you will be labelling it yourself.