YouTube Video Summarizer — Summary, Key Points and Timestamps

YouTube video summarizer for any link: paste a video URL and get back a summary in the style you choose, the key points, and a timestamped link to the moment each point is made. Built for researchers, students, editors and analysts who need to know what a video says before they give it an hour of their day.

A full link, a youtu.be link, a Shorts link or a bare video id all work.

Why not just read the transcript

Because a transcript is not a summary. An hour of speech is thousands of words of false starts, filler and "so basically", and nobody reads that to find out whether the video was worth watching. Transcript downloaders hand you all of it. This hands you the answer, plus the exact seconds the answer came from.

Four styles, for four different jobs

• Brief — three or four sentences. For triage: is this video worth my time?
• Detailed — four short paragraphs covering the argument, the evidence and the conclusion. For when you are not going to watch it and need to know what it said.
• Bullet points — six to ten complete sentences, one claim each. The best one to paste into notes.
• Chapters — the video broken into its natural sections. For long talks and interviews.

Timestamps are matched, not invented

This is the part worth understanding, because it is where these tools usually quietly lie.

A model asked to produce timestamps will produce convincing ones that are not in the video. So no timestamp here comes from the model. The key points it writes are matched back against the transcript's own segments, and the timestamp you get is the start of the segment those words actually came from.

A point that cannot be matched to a segment gets no timestamp rather than a guess. One shared word does not count as a match, because "video" appearing in both is a coincidence, not a location. Each matched point also carries a url that opens the video at that second, so a reviewer can check the claim in one click.

Long videos are summarised in full

A transcript too long for one pass is split, each piece is summarised, and those summaries are summarised again. transcript_chunks tells you how many passes it took.

The alternative, which is what a single call does when overfed, is to silently drop the end. On a two-hour talk that means the conclusion is missing and nothing in the output says so.

Videos without captions

A large share of YouTube has no captions at all. Those come back with summarised: false and an error saying so, and they are not charged — the per-video price only applies to videos that produced a summary.

Prefer human captions is on by default. Auto-generated captions mishear names, numbers and technical terms, and those mistakes then get repeated in the summary as if they were facts. When a human-written track exists it is used.

Summary language

Summary language writes the summary in whatever language you name, regardless of what the video is in. Leave it empty to follow the transcript. This is separate from Transcript languages, which chooses which caption track to read. A Dutch summary of an English lecture is one field away.

Input

The form on the Store page is the whole input; the same fields work as JSON when you call the Actor through the Apify API.

Output of the YouTube video summarizer

One dataset row per video, in input order, with the run counters in the key-value store under summary and run-level errors under errors.

A real row, trimmed only by dropping the middle eleven of the thirteen key points (first and last kept), from a 18-minute lecture with 3,357 transcript words:

{
  "video_id": "aircAruvnKk",
  "video_url": "https://www.youtube.com/watch?v=aircAruvnKk",
  "language": "en",
  "is_generated": false,
  "duration_seconds": 1105.64,
  "word_count": 3357,
  "summary_style": "brief",
  "summary": "This video explains the basic structure of a \"plain vanilla\" neural network designed to recognize handwritten digits. It describes how a network uses layers of neurons—consisting of an input layer for pixels, hidden layers for detecting edges and patterns, and an output layer for digit classification—to process information. The author concludes that the network functions as a complex mathematical function where weights and biases act as \"knobs and dials\" that can be tuned to recognize specific patterns.",
  "key_points": [
    {"point": "Neural networks are inspired by the brain and can be used to recognize handwritten digits.",
     "seconds": 163.38, "timestamp": "2:43",
     "url": "https://www.youtube.com/watch?v=aircAruvnKk&t=163s"},
    {"point": "Modern networks often use ReLU (Rectified Linear Unit) instead of sigmoid functions because it is easier to train.",
     "seconds": 1059.4, "timestamp": "17:39",
     "url": "https://www.youtube.com/watch?v=aircAruvnKk&t=1059s"}
  ],
  "chapters": null,
  "position": 1,
  "key_point_count": 13,
  "transcript_chunks": 1,
  "summarised": true,
  "error": null
}

The matching run summary in the key-value store reads

{"videos": 1, "summarised": 1, "withoutTranscript": 0, "totalWords": 3357, "withTimestamps": 1, "style": "brief", "fetchedInSeconds": 6.83, "errorCount": 0}

A video with captions switched off comes back as a row with summarised: false and error explaining it, and the run still succeeds — a failed video is data, not a crashed job. A run whose links are all blank produces no rows and succeeds as well.

Use cases

• Deciding what to watch. Run a week of subscribed videos through brief and keep the four that earned an hour.
• Literature and interviews. Pull key_points with their timestamps, then open only the cited seconds. The claim and its location arrive together.
• Meeting and lecture notes. Bullet points pastes straight into a doc, and outputLanguage writes it in the language your team works in.
• Feeding a model. One row per video gives a clean, structured context block — summary, points, timestamps, and the transcript only when you turn includeTranscript on.
• Batch triage. List dozens of links, watch transcript_chunks and key_point_count to see which videos were long enough to be worth a second pass.

How it compares

• Transcript downloaders (including our own free YouTube transcript Actor) give you the raw text: accurate, complete, unread. This gives you the answer and the seconds it came from.
• Browser extensions and one-off web tools such as NoteGPT, MindGrasp, Decopy and Noiz summarize the video you are looking at, in a tab, by hand. Useful once; they do not take a list of fifty links, return JSON you can join on, or tell you which caption track was used.
• Asking a chat model for a summary is the "youtube summary with chatgpt" route: you paste the transcript yourself, you hit the context limit on long videos, and you get timestamps the model invented. Here the transcript is fetched for you, long ones are split, and every timestamp is matched against the transcript's own segments.
• Generic AI summarizers summarise text you hand them. This one reads YouTube's captions, prefers human-written tracks, and refuses to invent a second that is not in the video.

FAQ

Does it work on videos without subtitles?

No, and it says so instead of guessing. A video with captions disabled comes back with summarised: false and an error naming the video, the run succeeds, and you are not charged for that video. Check whether your videos have captions before you promise a colleague a summary.

Can I summarize a whole playlist or a channel?

You can paste as many links as you like — one video per line, up to the limit your plan allows. The Actor reads captions, so a playlist becomes one row per video, in the order you pasted them. Building the link list is a job for a search or channel scraper; this Actor does the reading.

Are the timestamps reliable?

They are matched, not generated. Each key point is compared against the transcript's own timed segments, and the timestamp is the start of the segment the words came from. A point with no good match gets no timestamp at all rather than a plausible-looking wrong one. Every point also carries a url that opens the video at that second, so verification is one click.

Which summary style should I use?

Brief for triage, detailed when you will not watch the video at all, bullets for notes you will paste somewhere, chapters for talks and interviews where the structure is the point. The style is a field, so you can run the same links twice and keep both.

Can I get the summary in another language?

Yes. Summary language writes the summary in the language you name regardless of what the video is in, while Transcript languages picks which caption track to read. Summary in Spanish from an English lecture is one field.

How do I run it without code?

The input form on this page is the whole input: paste links, pick a style, hit Run. For anything repeatable, schedule the Actor or call it through the Apify API with the same JSON fields listed under Input.

Pricing

Pay per event, no subscription and no minimum. Two events make up the bill:

• Per-run start fee — charged once per run, $0.002, whether the run returns one row or fifty.
• Video summary — $0.02 per video actually summarised. Videos that came back without captions are not charged, and neither is anything you did not get.

A run over ten videos therefore costs $0.002 plus ten times $0.02, and a run that finds nothing costs the start fee alone. Users on a free Apify plan are delivered at most ten videos per run; paying users get the whole list. Prices are the ones on the Store page at the time of writing.

Errors

The two are deliberately separate: one is YouTube's side, the other is ours, and only one of them is worth retrying.

Related actors

free-youtube-transcript if you want the raw transcript and no summary, youtube-scraper for search results, and youtube-comments for what people said underneath.