The words, not a shorter version of them

YouTube Transcript: Every Word, With Timestamps

Sometimes you do not want the gist. You want the sentence someone actually said, so you can search it, quote it, check it, or paste it somewhere. That is a transcript, and every video run through this site produces one.

Paste a video link →

Get the transcript, then switch to the summary

TL;DR

Building a GPT from Scratch: Let's Reproduce GPT-2 (and ChatGPT?)

We build a character-level Transformer language model from scratch using PyTorch, training on Tiny Shakespeare, and explain the architecture of GPT.

Andrej Karpathy · 1h 56m · about 3 min to read · about 1h 53m saved · 8 key points · timestamped

The lecture begins with a demonstration of ChatGPT and explains that its underlying architecture is the Transformer, introduced in the 2017 paper 'Attention is All You Need'. The goal is to build a simpler, character-level version of GPT using PyTorch, trained on the Tiny Shakespeare dataset. The process starts with tokenization, where text is converted to integer sequences using a vocabulary of 65 characters. Data is split into training and validation sets, and batches of fixed-length chunks are created for training. A bigram language model is implemented first, which predicts the next character based only on the current character. This baseline achieves a validation loss of about 2.5. The core innovation is self-attention: tokens communicate by emitting query, key, and value vectors. The attention scores are computed as dot products between queries and keys, masked to prevent future tokens from attending to past ones, then softmaxed and used to aggregate values. This is efficiently implemented using matrix multiplication with a lower triangular mask. Multi-head attention runs multiple self-attention heads in parallel, concatenating their outputs. A feed-forward network (a simple MLP) is added after attention for per-token computation. Residual connections and layer normalization are introduced to stabilize training of deeper networks. The model is scaled up to 384 embedding dimension, 6 heads, 6 layers, and dropout of 0.2, achieving a validation loss of 1.48 on Tiny Shakespeare. The resulting generated text is nonsensical but resembles Shakespearean style. The lecture concludes by contrasting the decoder-only Transformer (used in GPT) with the encoder-decoder architecture from the original paper. The pre-trained model is a document completer; to turn it into a helpful assistant like ChatGPT, additional fine-tuning stages are needed, including supervised fine-tuning and reinforcement learning from human feedback (RLHF). The full code is available in the nanoGPT repository.

Transcript is the default view on this page. Switch to Full Summary when you want the shorter, structured version of the same speech.

Getting the text out of a video

1Paste a video link in the box aboveAny video anyone can open — a normal watch link, a youtu.be short link, a Short, or a finished livestream.
2Wait for the run to finishThe caption track is fetched before anything else is written, so the text exists whether or not you care about the notes that come with it.
3Read the Transcript tab hereThe transcript opens automatically when processing finishes. Copy the text, follow a timestamp to the video, or switch to Full Summary for the key ideas.

Captions, transcripts and subtitles are three different things

People use the three words interchangeably, and then wonder why the file they downloaded is not the thing they wanted.

Captions

The timed text track YouTube paints over the picture, three or four words at a time. Either uploaded by the channel or produced by speech recognition. Built to be read in sync with what you are watching.

A transcript

The same words lifted out of that sync and laid out as continuous text. One page you can scroll, search with Ctrl-F, and quote from, instead of a track that scrolls away while you read it.

Subtitles

Usually a translation, for people who do not speak the language being spoken. Auto-translated subtitles are generated from the caption track, so they inherit whatever the captions got wrong.

One consequence is worth stating plainly: what you get here is the caption track, warts and all. Speech recognition drops punctuation, runs sentences together and spells unfamiliar names by ear. Those errors are repaired in the notes written alongside, where inferring from context is legitimate — never in the transcript, where the whole point is that the words are the ones that were spoken.

Why it is grouped into blocks

A caption line is three or four words long. A two-hour video is several thousand of them, and a page built one line per row reads like a receipt. The lines are therefore regrouped into blocks of about 45 seconds on the page, and 30 seconds in the downloads, each keeping the start time of its first line.

Every block time is a link into the video at that second, so a sentence that looks too good to be true takes one click to verify against the recording.

Times are always written HH:MM:SS, including the hours on a ten-minute clip. The shorter MM:SS form breaks down on long recordings — a four-hour course produces markers like 223:58, which no player and no parser will accept. That was measured before it was fixed, which is why the zero-padded hours are there.

Copying it, and taking it with you

Expand the transcript, select it, copy. It is ordinary selectable text in the page — not an image, not a canvas, not locked behind an account. Reading the whole thing and copying it out of the page costs nothing and asks for nothing.

Downloading it as a file is the part that is paid for, and three formats exist:

Markdown — front matter carrying the title, channel, duration, date and source URL, then the notes and the whole transcript, with every timestamp already a clickable link. Drops into Obsidian or anything else that reads .md.
Plain text — the same content with no markup, for pasting into an email, a ticket, or another tool.
Anki — a flashcard deck built from the key points rather than the transcript, since nobody wants to revise four hours of speech one caption line at a time.

Where the line sits between the free and paid halves is set out on the pricing page. The short version: everything is readable on the page for free, and files cost money.

Which language you get back

Whichever one was spoken. The track that was used is recorded with the text — a plain code such as en or de for a caption track someone uploaded, and a machine-generated marker for one produced by speech recognition, which is a useful thing to know before quoting from it.

Nothing is translated on the way through. A talk given in Spanish comes back in Spanish, and the notes written from it are written in Spanish too. If you need English out of a Portuguese lecture, that is a translation step you would have to do yourself for now.

Transcript questions

Is the transcript word for word?

For the words, yes: it is the caption track exactly as the provider returns it, with the short caption lines joined into readable blocks and the start time of each block preserved. What it is not is tidied up. Machine-generated captions arrive without punctuation, without paragraph breaks and with names spelled by ear, and that is left alone here — editing the wording would quietly turn a quote into a paraphrase.

What happens when a video has no captions?

Two providers are tried in turn. The second one can transcribe the audio itself, which covers videos with no caption track at all. If neither can produce speech text, the tool reports that no transcript is available for that video rather than writing something plausible.

Can I get the transcript of a private or members-only video?

No. Only videos any signed-out visitor can open. If a link needs an account, a purchase, or a membership, the transcript providers cannot reach it either — and a browser extension running inside your own logged-in session is the tool that fits that case.

Can I have it translated?

Not yet. Text comes back in the language that was spoken, and the notes written from it are written in that same language. Translation is a separate step that is not built.

How long a video can it handle?

The longest run tested end to end is a four-hour programming course whose transcript is over 250 KB of plain text. Length costs waiting time on the first run, not completeness — the text goes to the end of the recording.

YouTube already has a transcript button. Why use this?

It does, on desktop: “Show transcript” under the video, when captions are enabled. That panel is genuinely good for following along beside the player, and if that is all you need, use it — it is closer to hand. What it does not give you is a page you can download, a copy that survives the video being taken down, per-block deep links, or notes written from the same text. That is the gap this fills.

Is the transcript stored?

Yes. Fetching one costs money and a video can disappear, so a fetched transcript is cached and reused instead of bought twice. The privacy page lists exactly what is kept.

Run a video — the box that takes a link, and examples of what comes back.
Choosing a summarizer — how a web tool, a browser extension and a chatbot differ when you want notes rather than the raw text.
About this site — the pipeline behind it, and the measured numbers it publishes.
Privacy — what is kept once a transcript has been fetched.