On-device video transcription — nothing leaves your browser

Captions or the raw file

The video tool takes one dropzone. Already have captions? Drop in a .vtt or .srt file, or paste a transcript. Only have the raw footage? Novus Learn can transcribe an .mp4 or .mp3 for you — entirely in the browser.

If you have captions, use them. They are usually more accurate than anything a small model will produce, they are already timed, and they skip the download and the wait entirely. On-device transcription exists for the case where captions don't exist.

That case is common enough to be worth solving properly. A lecture someone recorded on a phone, a conference talk that was never captioned, a voice memo of a meeting you need to be accurate about — none of these come with a transcript, and the usual fix is to upload them to a service that will happily keep a copy. For recordings involving other people who did not agree to that, this is not a small detail.

The model runs on your device

On-device transcription uses a compact Whisper speech model, quantised to keep it small, and served from our own origin rather than a third-party CDN. You download it once — explicitly, as its own step, at roughly eighty megabytes — and it is then cached for later use.

Your audio and video are never uploaded. The page only ever talks to its own origin, and the audio is decoded to the mono 16 kHz stream the model expects and processed in overlapping thirty-second windows locally. Long recordings simply take longer; they do not get sent anywhere to be handled elsewhere.

Why the download is opt-in

Eighty megabytes is not a rounding error on a phone plan, and a model that appears on your device without being asked for is a decision made on your behalf. So it isn't: before anything is fetched, Novus Learn checks whether your device can plausibly run it and asks you.

The storage check is a caution rather than a gate. If free space looks tight you'll be told, and you can still proceed if you know better than the estimate. Nothing about the video tool is blocked until you opt in — the captions path works with no download at all.

What to expect from it

Realistic expectations make this feature much more useful, so: the model is a small English one, chosen because it has to run on a laptop inside a browser tab. It is not the largest or most accurate speech model available, and it was never going to be.

Speed depends heavily on the machine. A recent laptop handles a short clip comfortably; an older or thermally-limited device will take noticeably longer, and a phone longer still. The work happens in overlapping windows across the recording, so progress is steady rather than sudden, and a long lecture is a background task rather than an instant one.

Audio quality matters more than length. A clear single speaker with a decent microphone transcribes well. A recording made from the back of a room, or a call with crosstalk, produces a transcript that needs real editing — at which point pasting in existing captions, if any exist, is simply the better move.

Review the transcript before you study it

Speech recognition is not perfect and this model is deliberately a small one. Review and correction is a required step, not a suggestion: you read the transcript and fix it before it becomes study material, so the resulting flashcards and notes are grounded in what was actually said.

Expect the usual failure modes and check for them specifically:

• Proper nouns, names and place names — the most common errors, and the ones most likely to end up on a card.
• Technical terms and acronyms, especially in a field the model has little exposure to.
• Numbers and units, which are easy to mishear and expensive to get wrong.
• Overlapping speakers and heavy background noise, where whole phrases can be lost.
• Accents and languages other than English — the shipped model is an English-only one.

From transcript to study set

Once you're happy with the transcript, it feeds the same pipeline as any other source: summary, notes, flashcards, quiz, and the rest — each tied back to the moment in the transcript it came from. A lecture becomes something you can test yourself on rather than something you have to rewatch.

Because the transcript is the source, the evidence links point at your own corrected text. That is the part that makes the review step worth doing properly: whatever you leave in becomes the thing the study set is grounded in.

It's also the reason this tool earns its place rather than being a novelty. A recorded lecture is one of the least efficient study materials there is — linear, unsearchable, and impossible to skim. Turning it into text you can search, cite and test yourself against changes what it's for, and doing that without sending someone's classroom recording to a server is the part we cared about getting right.