About Summarize YouTube Videos

This is a small, independently run tool that turns a public YouTube video into something readable: the claims it makes, chapters that point at the moments they were made, and the full transcript underneath both. It has one input — a link — and nothing to sign into before you read what comes back.

How it works, in four steps

  1. 1A transcript provider supplies the speechTwo are wired up. The primary is fast and reads the caption track; the fallback can transcribe the audio itself, which covers videos that have no caption track at all. Fetching directly from YouTube is not an option and was not a preference — that path was measured and dies after roughly 20–25 requests from any single address.
  2. 2The text is handed to a language model with its timings attachedEvery line arrives prefixed with its own HH:MM:SS marker and the video's total duration. This is the part most implementations skip, and it is why their chapters bunch up at the start: a model given a plain wall of text has no way to know when anything was said, so it guesses.
  3. 3The model returns structure, not proseThe response is a single JSON object — key points, a TL;DR, flowing summary, chapters — parsed progressively as it streams, so points appear one at a time instead of behind a spinner. Timestamps must be copied from the transcript verbatim; the model chooses which moments matter, never what time they happened.
  4. 4The result is cached, and the transcript is cached separatelyTwo tables, because they expire for different reasons. Rewriting the prompt makes summaries stale while the transcript underneath stays perfectly good, and a transcript cannot always be bought again — a deleted video takes its captions with it.

The providers are named on the privacy page rather than kept vague, because knowing whose servers a link passes through is a reasonable thing to want before pasting one.

Numbers we publish, because they are measured

Every tool in this category describes itself as fast, accurate and free. None of those words can be checked. These can — they come from instrumented runs, not from a marketing meeting, and they are the ones we would want to see on somebody else’s site.

$0.012

to summarise a four-hour course

The expensive end. Model and transcript spend for the longest thing tested end to end; a short talk is a fraction of a cent.

8ms · $0

for a repeat of a video already run

The second person to ask for a video is served from cache. It costs nothing and returns before the page has finished laying out.

87–94%

of runtime covered by chapters

Measured span from the first chapter to the last, against the video's real duration. The failure mode this catches is chapters that stop ten minutes into a two-hour talk.

The cost figures matter for a reason that has nothing to do with transparency theatre. When a service is free, somebody is paying, and how much they are paying determines whether it is still there in a year. A tenth of a cent a video is a hobby that can absorb a bad week; a dollar a video is a countdown. Publishing the real range lets you make that judgement instead of guessing — and the same numbers appear on the pricing page, where they explain why file exports are the paid half and reading is not.

The coverage figure is there because it is the number this pipeline was hardest to get right, and the one nobody else reports. It went up sharply once the transcript was fed to the model with its total duration attached; before that, chapters on a two-hour lecture covered 11% of it. That fix is a line in a prompt, and the reason it was found is that the number was being watched.

Numbers on this page are re-measured when the pipeline changes. If one of them looks wrong to you, the transcript and the timestamps are published on every result page, so it is checkable rather than arguable.

What it does not do

Summaries can be wrong

Speech recognition mishears, and a model reading a mis-transcribed sentence will summarise the mistake with total confidence. This is why the transcript ships alongside every summary rather than being hidden: anything surprising can be checked against the words in one click.

Public videos only

If a link needs an account, a purchase or a membership, the transcript providers cannot see it, and neither can this.

It is early, and says so

Reading a result needs no account. Passwordless accounts, subscription billing, exports and playlist batches exist, while a registered business identity, translation and follow-up questions do not. The privacy and terms pages state those gaps rather than papering over them.

Not affiliated with YouTube or Google

It reads public videos through third-party transcript providers, like any other independent tool in this category.

Who builds it

It is built and operated independently, not by a company with a support department. That is visible in the product: the free tier is not crippled to sell the paid one — the entire result, transcript included, is readable without paying — and nothing on the page interrupts you to collect an address. It is also visible in the failure modes — when something breaks, it is fixed by whoever is at the keyboard, and mail sent to the contact address lands with whoever that is rather than in a ticket queue.

Decisions here are argued from measurements, and the measurements are kept even when they are unflattering. The 11% chapter coverage mentioned above was our own bug, found by instrumenting the thing rather than by a user complaining.

Corrections are welcome, including to this page. Get in touch.

Elsewhere