Lecture to Anki: Convert the Notes, or the Recording Itself?

The search "lecture to anki" is typed by someone holding the wrong artifact. There is a recording on the phone — the lecture captured whole — and an exam sitting nine days out. What ranked for the phrase when I checked is a row of small converter sites: ankify, AnkiGPT, 2anki, StudyCardsAI. Every one of them eats text. Notes, PDFs, slide decks; paste, convert, download. They answer a question one step downstream of yours, because the thing in your hand is not text at all. It is fifty minutes of speech.

The audio slot is not entirely vacant — AnkiDecks, at this writing, takes an audio file and hands back an .apkg, five minutes of audio on its free tier. Five minutes is a vocab clip. A lecture is a different quantity of problem, and "convert all of it" is the wrong answer to that problem anyway, for reasons worth a round of arithmetic.

Usual disclosure up front: I build Piccard, a flashcard tool whose upload route is one of the three assessed below. Grade my homework accordingly.

Why the recording is a different job

Notes are already selection. Whatever is in your notes, the professor said it and you judged it worth writing down; a converter inherits that filter for free. A recording has no filter. Run the numbers once: fifty minutes of lecturing at 130 to 150 words a minute is around seven thousand words. Spoken sentences run short, maybe fifteen words each, so call it four hundred and change sentences per lecture. Feed that transcript to any converter and you get four hundred cards from one session — and a twelve-week course meeting twice a week heads toward ten thousand. Nobody reviews ten thousand cards. The deck dies in week three under its own backlog.

So the real job was never conversion; it was selection. Which of those four hundred sentences carry something worth a card is exactly the question a transcript cannot answer and a converter does not ask. The general filter has its own write-up: Which English Words Are Worth Learning? A Three-Part Filter. The short version for a lecture hall: a phrase earns a card when you will meet it again and do not yet own it. The professor's verbal filler fails both tests, and a transcript preserves all of it.

Three routes from a recording to cards

Play and type. It works and it costs the evening, plus a flaw that hides until exam week: the sentences you catch while keeping up are the ones you nearly knew. The term that slid past without a flinch needed the card more, and it is precisely the one absent from your notes. Pausing and replaying fixes this at the price of making one lecture a three-hour job.

Transcript plus converter. This is the route the ranking sites serve, one step removed: run the recording through a transcription service — the free tier of Whisper-based tools) handles a lecture adequately — then paste the text into a notes-to-deck tool. If what you have is actual notes, that is the right tool for them. But a raw lecture transcript is not notes. It keeps every "so", every "right?", every second pass at the same explanation, and the converter will faithfully deck all of it. You solve the transcription problem and inherit the selection problem untouched.

Upload, score, confirm. Transcription and ranking in one pass; the cards exist once you confirm the selection. This is the route Piccard takes, and the next section is what it does, as built today.

What the upload path does, feature by feature

The upload. An audio file up to 90 minutes fits in one piece — a standard lecture, whole — and the duration cap is a gate, not a suggestion: a file over it is rejected before upload with the reason named. Any file over 100MB is rejected the same way, which in practice bounds video at about a quarter hour of mp4 — pull the audio off a session video first. The audio must be in the language you are learning, enforced rather than hoped for: on this route that means English, so a Japanese lecture recording — someone else's target language — is rejected with a named reason, as is a silent file, instead of a junk transcript. The gate matters because everything downstream is only as good as the transcript.

The score. Every phrase in the transcript comes back with a 0-to-100 worth-learning score, and the free-text learning goal you have set in the app is weighted into it — the same lecture scores differently for "I need to follow pharmacology lectures" than for "general academic English". Hover a score and the reason rides along with it, so the ranking is an argument you can inspect. Phrases at 50 and above arrive pre-checked, and the list is sorted so those sit at the top: the triage is a scroll, not a search.

The confirm. One confirm turns at most 20 phrases into cards. The cap tracks what a confirm costs the model — a few hundred generated tokens per card — and nothing else. A dense lecture where the scorer surfaced forty keepers is two confirms.

The cards and the schedule. Each card carries the term with its meaning, an example sentence, and a pronunciation field. Review runs FSRS-6 at a 0.9 target retention with no settings page to get wrong — FSRS is the scheduler family Anki itself ships now, so the engine was never the disagreement — and a card that keeps surviving graduates at 180 days. Once it exists, the lecture deck is not special; it is cards on the same forgetting curve as everything else, resurfaced by spaced repetition just before each drop.

If what you have is notes. The paste route lives in the same app: text, PDF, .docx, .txt, or .md comes back as candidates ranked high, medium, or low priority — a different ranking from the audio path's 0-to-100, computed client-side, with the highs pre-selected. If the department hands out slides, that is the cheaper door into the same review queue.

The limits, plainly. There is no .apkg export; cards review inside the app, and if your entire study life runs through Anki's deck list, this is not your tool for this job. The audio language is a hard boundary — uploads transcribe the language you are learning, and on this route that is English. A Japanese lecture recording is accepted only for someone learning Japanese; here it is rejected with a named reason, and the rejection keeps a 2-energy detection fee rather than costing nothing. And the Chrome-extension half of the product is a different machine: bilingual subtitles on YouTube videos that have captions, tap a line to get candidates. It needs a caption track; it does not transcribe files.

The semester arithmetic

1. One lecture, one confirm: 20 cards is a normal take, and a dense lecture runs two. That is the whole per-lecture cost; the selection work is the scroll-and-uncheck.
2. A twelve-week course at two lectures a week, keeping 20 cards a lecture, is 480 cards. Spread across the term that is about six new cards a day, a load a 0.9-retention schedule holds without heroics.
3. A week before the exam, stop adding. The lecturer's own slide deck is the coverage audit: walk its term list once, pull the items you cannot produce on demand, and let the queue drain.

The short version

The converters ranking for "lecture to anki" are the right tools for the notes you took and the wrong ones for the recording you made. Text in, deck out is a solved problem, and a lecture is not text: the recording needs transcription, the transcript needs triage against your goal, and the survivors need a schedule. The last part was always the actual product. The deck built from your own material is also the one nobody can hand you — the same argument from the Duolingo side of the fence is its own post — while the downloadable kind of deck answers a different question, a language's starter band rather than your own lecture hall, and the Japanese deck comparison maps that category.

The free tier is 150 energy — transcription meters from the same balance at 1.5 a minute, so a fifty-minute lecture is about 75 energy to transcribe plus 1 per card, roughly one and a half lecture round-trips inside the grant — then prepaid top-ups at $5 / $10 / $20; no subscription, and topped-up energy never expires. To start: Piccard on the Chrome Web Store.

All guides: Piccard Blog.