Novus Learn
Break down a video in Novus Learn, with on-device transcription
The /video surface turns a lecture, talk, or tutorial into study material — working from captions where they exist and transcribing on your own device where they do not.
A two-hour lecture contains maybe twenty minutes of material you need, and no way to find it except watching the whole thing at 1.5×. Video is the worst format to study from and, increasingly, the format everything ships in first.
The /video surface fixes the retrieval problem without breaking the sourcing rule. It builds a text source from the video — captions if they exist, on-device transcription if they do not — and then extracts claims from that text the same way it would from an article.
Contents
Two ways to finish
Captioned video
Fastest path: the existing captions become the source.
No captions
Transcription runs on your machine, so the audio never leaves it.
- 1
1. Start from captions when they exist
Open /video and point it at the material. If captions are available, Novus Learn uses them: they are already timed, already text, and already what the speaker said. This is the fast path, and for a well-captioned conference talk or lecture it is close to instant.
Auto-generated captions are worth a moment of scepticism, though. Speech recognition mangles proper nouns, technical vocabulary, and numbers more than it mangles ordinary prose — which is unfortunate, because proper nouns, technical vocabulary, and numbers are usually the parts you wanted.
- 2
2. Fall back to on-device transcription
Where there are no captions, Novus Learn transcribes the audio on your device. The model loads in the browser and runs locally, which means the audio does not travel to a transcription service and no third party ends up with a copy of the recording.
The first run pays a model-download cost, and transcription is real work for your machine — a long video on a light laptop takes a while. That is the honest trade for keeping the audio local, and it is the same trade the rest of the portfolio makes.
- Transcription runs in your browser; the audio never leaves the device.
- First run downloads the speech model, then it is cached.
- Long videos on light hardware take real time — this is compute, not a queue.
- 3
3. Read the transcript before you trust the claims
Whether it came from captions or transcription, the transcript is now the source, and everything downstream inherits its errors. Skim it for the terms that matter to you. A misheard technical term will propagate into a flashcard and quietly teach you something wrong, and that is far more likely than the extraction step introducing an error of its own.
This is the video-specific version of the habit that runs through the whole app: check the source layer, because the artifacts are only ever as good as what they were built from.
- 4
4. Extract claims and find the twenty useful minutes
With the transcript in place, the claims list does the job that made you want a tool in the first place. Instead of scrubbing a timeline, you read the assertions the speaker actually made, and the ones worth your attention are visible at a glance.
Use the claims as an index back into the video. When a claim matters, go watch that part properly — tone, emphasis, and what a speaker chooses not to say are things a transcript flattens.
- 5
5. Build notes, cards, and a map from the talk
From here it is the standard set: summary, notes, flashcards, quiz, concept map. For a lecture, notes and cards are usually the pair that earn their keep; for a conference talk that is mostly argument, the concept map tends to expose the structure the speaker was building toward.
Flashcards export as a printable Markdown answer key with the source behind each card, which turns a one-off talk into revision material you can use months later without re-watching anything.
- 6
6. Combine it with the other input modes
Video rarely stands alone. A talk usually references a paper, a documentation page, or a repository, and Novus Learn can open all three: /paper for the article, /read for the page, /code to x-ray the codebase. Studying the talk alongside the thing it is about is a far better use of the tool than treating any one source as complete.
If you are teaching from it, /teachers turns the project into a printable lesson with objectives, a worksheet and answer key, and citations — so the class gets the material without watching two hours of video first.
The transcript is the source — check it first
Captions are the fast path; on-device transcription is the private one. Either way, skim the transcript for the terms that matter before trusting anything built on it, because a misheard technical term propagates silently into cards. Use the claims list as an index back into the video rather than as a replacement for watching the parts that count.
Frequently asked questions
Quick answers to common questions about this topic.
Does my video get uploaded?
No. Where captions exist they are used directly; where they do not, transcription runs on your own device in the browser. The audio does not travel to a transcription service.
Why is the first transcription slow?
The speech model downloads on first use and then caches. After that, the cost is your machine doing the actual transcription work, which scales with the length of the video and the hardware you are on.
How accurate are auto-generated captions?
Good on ordinary prose, unreliable on proper nouns, technical vocabulary, and numbers — which are often the parts you care about. Skim the transcript for your key terms before relying on anything built from it.
Can I still get flashcards from a video?
Yes. Once the transcript exists it behaves like any other source, so you get the full set: summary, notes, flashcards, quiz, and concept map, each cited back to the transcript.