Flashcards from YouTube and TikTok videos: how we turn subtitles into cards with Supadata
Paste a YouTube link and get a flashcard deck from its captions. How it works, what it costs, why native video helps language learners, and the Supadata build.
Paste a YouTube link into Memor More and the app makes a flashcard deck from what the speaker says. The server reads the video's captions through Supadata, a transcript API, and hands the text to the same AI prompt that turns a PDF into cards. Nothing is downloaded, and a video that already has captions costs us one credit.
Credit where it is due: Supadata is the reason this feature took days rather than months. One endpoint, pricing you can work out on a napkin, and documentation that answered our questions before we had asked them. More on that below.
YouTube links come first, as part of the next Memor More update for iPhone, iPad and Mac. TikTok links follow. This article covers why we wanted video as a source at all, how the feature works from the learner's side, what a real video turned into, and the engineering Supadata saved us. Product details were checked on 9 November 2026.
Why we wanted video as a source
A textbook teaches you la cuenta, por favor. A Madrid vlogger says ¿nos traes la cuenta cuando puedas? while the waiter is already walking away. The second version is the one you'll hear, and the one you'll remember, because you watched it happen to someone you'd been following for twenty minutes.
That is the case for native content: vlogs, street interviews, cooking channels, gaming streams, ten-second TikToks. It carries vocabulary that textbooks avoid, in the register people actually use, with the speaker's face and the situation attached. The memory research is careful here and so are we. Reviews of emotional memory, such as LaBar and Cabeza's 2006 survey in Nature Reviews Neuroscience, describe how emotionally charged events get privileged treatment at encoding and consolidation. Those are laboratory findings about events, not a measured result for vocabulary learned from vlogs. What we can say is that a word with a scene attached gives you a second route back to it, and that this is a common reason learners prefer video to word lists.
The evidence that video itself teaches words is more direct:
- Watching full-length TV teaches vocabulary incidentally. Peters and Webb (2018) had Dutch-speaking English learners watch one full-length programme and found learning at the level of meaning recall and meaning recognition. Learning depended on how often a word occurred, on prior vocabulary knowledge and on cognateness.
- Captions help, and same-language captions help most. Reynolds, Cui, Kao and Thomas (2022) reviewed 34 studies and found that viewing captioned or subtitled video had a positive effect on vocabulary acquisition, with intralingual captions (the video's own language) showing the largest effect. They found no significant difference between intentional and incidental viewing conditions. The earlier meta-analysis by Montero Perez, Van Den Noortgate and Desmet (2013) found that captions helped both listening comprehension and vocabulary learning, and that learner proficiency did not change the effect.
- Purpose matters. A 2024 survey of 322 Chinese university students in ReCALL found that informal online English use predicted receptive vocabulary, but the link ran mostly through social and entertainment use rather than looking things up, and the purposes together explained only 15 to 23 percent of the variance. Watching is not a plan on its own.
The Peters and Webb frequency finding is the gap this feature fills. A vlogger uses a useful phrase once and moves on. Spaced repetition exists for words you met once and would otherwise lose, so the natural pairing is: watch for the context, then review the words the video gave you. The spaced repetition vocabulary schedule describes the daily routine that follows. For the general case for the method, see the language learners page.
Not only languages
YouTube is also where a lot of serious teaching lives. A university lecture series on the Roman Republic, a cardiologist walking through an ECG, a chemistry teacher explaining buffers for the third year running. The same feature works on all of them, because it starts from the caption text rather than from a language-learning template.
Learners preparing for a specific exam are the obvious case. Exam channels publish exactly the material a flashcard deck is made of, usually definitions and worked examples with the exceptions pointed out.
| Exam family | Examples | What a video deck tends to hold |
|---|---|---|
| Language certificates | IELTS, TOEFL, Cambridge B2 First and C1 Advanced, DELE, DELF and DALF, Goethe-Zertifikat, JLPT, HSK, TOPIK | Topic vocabulary, collocations, exam-task phrases |
| Medicine and nursing | USMLE Step 1, MCAT, NCLEX, PLAB | Mechanisms, drug classes, diagnostic criteria |
| School leaving | A-Level, GCSE, Abitur, AP, IB, PSLE | Definitions, formulas, dates, named processes |
| Professional and civil service | Bar exam, CFA Level I, oposiciones | Rules, thresholds, exceptions |
A video deck is a first draft. Lectures repeat themselves, speakers go off on tangents, and a 90-minute recording produces far more candidate facts than you want to review. The review pass described in how to turn notes into flashcards applies as written: delete cards that ask about the lecturer's anecdotes, split overloaded answers, and keep the cards that match the syllabus. Medical students will recognise the pattern from the USMLE workflow; exam candidates generally should read flashcards for exam revision first.
What the learner does
The feature lives where the other AI sources are: create a deck, choose AI Generation, then From a YouTube video. The screen carries a note that this is a test feature, because it is.
- Paste the link. Watch links, Shorts links and
youtu.belinks all work. The app parses the video id on the device and sends only that. - Check the preview. The server answers with the title, channel, thumbnail, length and the video's language, plus a Captions badge when the uploader added caption tracks. Most videos show no badge and still work, because YouTube's auto-generated captions are read at the next step.
- Pick the transcript language if the video offers several. The video's own language is first in the list.
- Choose what kind of cards you want. Definitions makes question-and-answer cards about the content. Vocabulary makes word cards and asks for your target language and CEFR level, so a B1 learner gets B1 words from an advanced vlog rather than the hardest ones.
- Set the number of cards and whether to keep the speaker's wording, then generate. Cards stream in as they are made, and you edit or delete them as with any generated deck.
The deck remembers where it came from. Its source note holds the video's title, link and transcript, so you can reopen the transcript later or jump back to the video. Deleting the transcript and link keeps the cards.
Some videos are refused before a generation is used: private videos, live streams and upcoming premieres, very long videos, and videos the provider cannot read because they are age-restricted, region-blocked or for members only. A video with no captions at all fails with "This video has no captions to make cards from," and the generation is refunded. So is a music video or a Short with almost nothing spoken in it, which is not worth a model call.
A video job uses one AI generation from your allowance, the same unit as a file or a topic. See pricing for the current free and Premium allowances.
One real video, end to end
We used Grant Sanderson's But what is a neural network? from the 3Blue1Brown channel, an 18-minute explainer, as the example video throughout development. Here is what the pipeline does with it, with the provider numbers from a request made on 9 November 2026.
| Step | What happened |
|---|---|
| Preview | YouTube's own API returns the title, channel, 1,120-second length and language en. No reject reason. |
| Transcript request | Supadata, mode=native, returns HTTP 200 and bills 1 credit (x-billable-requests: 1). |
| What came back | 286 caption lines totalling 18,145 characters, each with an offset and duration in milliseconds. 26 caption languages were available, from Arabic to Urdu. |
| First lines | "This is a 3." at 4.2 seconds; "It's sloppily written and rendered at an extremely low resolution of 28x28 pixels," at 6.1 seconds. |
| Prompt | An 18-minute transcript fits in a single prompt, so the model is called once with the title and channel as context. |
| Cards | Whatever number the learner asked for. Duplicate fronts are dropped. |
We have not pasted generated cards here, because they vary from run to run and a hand-picked set would say little about an ordinary result. Run it on this video yourself and compare.
A longer video changes one step. A two-hour lecture does not fit in one prompt, so the transcript is split into time windows that fall between caption lines, each window is generated in turn with a note about which minutes it covers, and the fronts are deduplicated across windows before the deck is finished. The timestamps Supadata returns with every caption line are what make that split clean.
What Supadata saved us
The AI prompt already existed for text uploads. The hard part was getting caption text reliably from a server.
A cloud server cannot read YouTube captions directly. YouTube blocks requests from datacenter IP ranges. We confirmed it the simple way: the same caption request succeeded from a home connection and failed from a cloud address with a sign-in challenge. Open-source scrapers hit the same wall, and cookies or proof-of-origin tokens are not a fix you want to maintain. Our card-generation model takes text, images and PDFs, not audio, so it could not transcribe the video either.
Supadata solves this with one endpoint. You send a video URL and an API key, and you get back either the caption segments or, for videos without captions, an AI-generated transcript. A few properties made it a good fit for a small team:
- The same call works for YouTube, TikTok, Instagram, X and Facebook. Our TikTok support reuses the YouTube transcript code almost line for line.
- Native captions first. We ask with
mode=native, which returns only existing captions, so we never pay for generation by accident. A generated transcript is a separate, explicit request that we keep behind a switch. - Timestamps come with the text. Each segment carries an offset and a duration, which is what we split long transcripts by.
- Long videos are asynchronous. Over 20 minutes the API answers 202 with a job id, and we poll it. Status checks are free.
- Pricing we could budget for. One credit per native transcript, 2 credits per minute of generated transcript, 1 credit for a "no transcript" answer. At the time of writing the free plan includes 100 credits a month, Basic $5 a month for 300, and Pro $17 for 3,000.
The integration itself took an afternoon. The rest of the time went into product work: previews, refunds, error messages a learner can act on, and the pieces below.
What we still had to build. A transcript API does not make a feature by itself. The things on our side that mattered:
- A shared cache. A fetched transcript is kept for a while under the language the learner asked for and the language it came in. The second person to paste the same lecture pays nothing, and the cache is cleared on a schedule so we never hold transcripts longer than we need.
- Remembering misses. Supadata bills a lookup that finds nothing, so a video that came back empty is not asked about again for a while. The preview shows the refusal instead, and no generation is spent.
- A rate gate. The entry-level plans allow a modest number of requests per second, and a burst of jobs started together was enough to trip it. Every Supadata call now goes through one queue with a small concurrency limit and backoff on a 429, so parallel jobs wait instead of failing.
- A monthly credit budget. When it is spent, new video jobs are refused, previews keep working, and the app shows the feature as temporarily unavailable rather than a limit nobody can act on.
- Preview without paying. For YouTube the preview comes from YouTube's own metadata API, which is free at our volume, so a learner can paste and reconsider as often as they like. TikTok has no equivalent API, so a TikTok preview is a one-credit metadata call that we cache and limit.
If you are weighing the same build: the API took an afternoon. The caching, the budgeting and the failure messages took the rest of the week.
Where it falls short
The transcript is only as good as the captions. Auto-generated captions mishear names and technical terms, and a card generated from a misheard word is wrong in a way that looks right. Check proper nouns and numbers in the generated deck against the video.
Caption text has no speaker labels and no slides. A lecture that says "as you can see here" produces a card about nothing. The Definitions mode handles this reasonably, but a diagram-heavy video is better served by a photo of the slide and the photo-to-flashcards path from version 1.11.
Because we only read existing captions, a video without any produces no cards. Many TikToks and some smaller YouTube channels fall into this group. We chose this deliberately: generated transcripts are priced per minute and we would rather refuse than surprise ourselves with a bill.
And this is not a listening exercise. The deck gives you the words; the video gave you the sound. Turn on Read Cards Aloud in Speech settings for the language and pick a Slow rate while the words are new, as described in the 1.11 release notes, and go back to the video when a card feels thin.
What's coming for TikTok
TikTok support is built and in testing. It differs from YouTube in a few ways that follow from the platform:
- The app sends the link as pasted, because short share links such as
vm.tiktok.comhide the video id. The server follows them. - The preview title is the post's caption with its trailing hashtags removed. Photo and carousel posts have no speech and are refused.
- There is no caption language picker, because TikTok captions come in the video's own language.
- Thumbnails are short-lived signed links, so the app never stores them with the deck.
Because the switch is on the server, TikTok links will start working without an app update once we are happy with it. If you want a deck to study in the meantime, the public deck library has language decks such as Spanish B1 emotions and relationships to try the review routine on.
If you want to try Supadata yourself, you can sign up through my referral link. It is a referral link, so I may receive a benefit if you sign up, at no cost to you.
Written by
Founder & developer of Memor More. I build iOS and Mac apps and write about the science of memory and learning. @Jerelii on X
