YouTube Transcript Scraper: captions, not speech-to-text

Four inputs, and a list of links it will refuse

The input surface is small and the listing is precise about its edges, which link forms work, which are rejected, and what happens to the ones in between.

• startUrls

  $0.001per transcript

  Only if it succeeds

  One video link per entry, as a plain string or as { "url": "…" }. Each is processed separately, so a duplicate link returns a duplicate row.

  https://www.youtube.com/watch?v=_AbFXuGDRTs

• A miss does not fail. It returns another track

  lang

  optionalno charge

  Check what came back

  An optional caption language code, en, es, de, tr, pt-BR. Leave it empty for the video’s default track.

  lang: "fr"

• maxItems

  ÷ $0.001caps the bill

  Run-wide

  Stops the run saving once this many rows exist. Leave it empty to process every link you supplied.

  100

• Unsupported links

  rejectednot charged

  watch and youtu.be only

  Shorts, live, embed, playlist and channel URLs are rejected outright, as are links from other sites.

  /shorts/… → rejected

Starting from a playlist or a channel? The listing points you at the Playlist and Channel Videos scrapers to collect the watch links first, then bring them here, which is exactly the pairing those two pages describe from the other end.

One video, a batch, or a batch in one language

Three runs. The billing model makes the second one unusually forgiving, and the third is where the one real subtlety in this actor lives.

{
  "startUrls": ["https://www.youtube.com/watch?v=_AbFXuGDRTs"],
  "lang": "en"
}

{
  "inputSource": "https://www.youtube.com/watch?v=_AbFXuGDRTs",
  "type": "transcript",
  "id": "_AbFXuGDRTs",
  "transcript": [
    { "text": "Hidden in this mountain is a $1 billion", "dur": "5.679", "startTime": "0:00" },
    { "text": "nuclear bunker. We are 2,000 ft", "dur": "4.8", "startTime": "0:03" }
  ]
}

{
  "startUrls": [
    "https://www.youtube.com/watch?v=_AbFXuGDRTs",
    "https://youtu.be/dQw4w9WgXcQ"
  ],
  "maxItems": 100
}

{
  "text": "nuclear bunker. We are 2,000 ft",
  "dur": "4.8",
  "startTime": "0:03"
}

{
  "startUrls": ["https://www.youtube.com/watch?v=_AbFXuGDRTs"],
  "lang": "fr",
  "maxItems": 100
}

{
  "selected": {
    "title": "English (auto-generated)",
    "languageCode": "en"
  },
  "availableLanguages": [
    { "title": "Vietnamese", "languageCode": "vi" }
  ]
}

One row per video, one array inside it

The shape is flat and small. Almost all of the payload sits in the transcript array, whose length is the length of the video rather than anything you control.

• 3

  The row

  Top level

  • inputSource, the link you sent
  • type, always "transcript"
  • id, the video ID

• 3

  Each caption line

  transcript[]

  • text
  • startTime, string, "0:03"
  • dur, string, "5.679"

• 2

  What came back

  selected

  • title, e.g. "English (auto-generated)"
  • languageCode
  • The field to trust

• 2

  Do not rely on

  availableLanguages

  • title, languageCode
  • One backend entry, not the video’s tracks

inputSource is more useful here than on most actors. Because each link is processed separately and a duplicate link returns a duplicate row, it is the field that lets you reconcile a returned set against the list you submitted and see exactly which links came back empty.

The limits the listing states plainly

This is the most carefully documented actor in the catalogue, it names its own failure code and warns against one of its own fields. All five below are its words, not inferences.

• 01

  It reads captions, it does not transcribe

  Uploaded or auto-generated tracks only. A video with captions turned off returns nothing, and speech-to-text is listed among the things this actor does not do.

• 02

  Only watch and youtu.be links

  youtube.com/watch, including m. and music. subdomains, and youtu.be. Shorts, live, embed, playlist and channel links are rejected, as are links from other sites.

• 03

  Skipped videos do not stop the batch or cost anything

  A video without captions is skipped with error code C003 and the rest of the run continues. You are charged only for transcripts actually extracted.

• 04

  lang is a request, not a guarantee

  If the video has no track in the language you asked for, a different track is returned, and billed. Check selected.languageCode and discard the mismatches yourself.

• 05

  availableLanguages is not a language list

  The listing says it carries a single entry from the transcript backend and does not enumerate the video’s caption tracks. Using it to decide what to request will mislead you.

Free accounts run in demo mode: 5 runs a month, 10 items each.

One event, and nothing else

The simplest billing in the catalogue. There is a dataset-item event, and the listing states plainly that this actor never charges it, so the transcript price is the entire cost, with no storage fee on top and nothing billed for a failure.

The last row is the one that changes how you work. Because failures are free, there is no penalty for throwing a whole channel’s links at it and keeping whatever comes back, which is not true of any other actor in the family.

Who needs the words out of the video?

Transcripts are the cheapest route from video to text, and text is what every search index, language model and analyst actually works on. The constraint is that the captions have to exist first.

• LLM and RAG pipelines

  A batch

  Turn a channel into a retrievable corpus, with timestamps kept so an answer can cite the moment

  $1.00 per 1,000 videos

• Search and indexing

  A batch, then re-run for new uploads

  Make a video library searchable by what was said rather than by what the title claims

  $0.10 per 100 videos

• Media monitoring

  A batch, scheduled

  Catch mentions inside long videos that no title or description would ever surface

  $0.01 per 10 videos

• Accessibility and localisation

  One language

  Collect the caption tracks that exist, checking selected.languageCode to see which languages a library really covers

  $0.10 per 100 checks

• Research corpora

  A batch, links from a playlist

  Build a labelled text set from a curated list, pairing this with the Playlist Scraper to get the links

  $10.00 per 10,000

From links to clean text

The actor is apidojo/youtube-transcript-scraper. Two things belong in every integration: check selected.languageCode, and strip the caption markup before anything treats the text as prose.

import re
from apify_client import ApifyClient

client = ApifyClient("YOUR_APIFY_TOKEN")

MARKUP = re.compile(r"\[[^\]]*\]|>>|\u200b")   # [Music], [ __ ], >>, zero-width

def clean(line):
    return re.sub(r"\s+", " ", MARKUP.sub(" ", line)).strip()

run = client.actor("apidojo/youtube-transcript-scraper").call(run_input={
    "startUrls": [
        "https://www.youtube.com/watch?v=_AbFXuGDRTs",
        "https://youtu.be/dQw4w9WgXcQ",
    ],
    "lang": "en",
})

for row in client.dataset(run.default_dataset_id).iterate_items():
    # lang is a request, not a filter: a miss returns another track, billed.
    got = row["selected"]["languageCode"]
    if got != "en":
        print(f"skipping {row['id']}: got {got}")
        continue

    text = " ".join(clean(l["text"]) for l in row["transcript"])
    seconds = sum(float(l["dur"]) for l in row["transcript"])   # dur is a string
    print(f"{row['id']}  {seconds/60:.1f} min  {len(text):,} chars")

import { ApifyClient } from 'apify-client';

const client = new ApifyClient({ token: process.env.APIFY_TOKEN });

const run = await client.actor('apidojo/youtube-transcript-scraper').call({
  startUrls: ['https://www.youtube.com/watch?v=_AbFXuGDRTs'],
  lang: 'en',
});

const { items } = await client.dataset(run.defaultDatasetId).listItems();

// Videos without captions are skipped and never billed, so a short result
// set is normal. inputSource reconciles what came back against what you sent.
const returned = new Set(items.map((r) => r.inputSource));
console.log(items.length, 'transcripts;', returned.size, 'of your links came back');

// startTime is "0:03", not a number, keep it for citations, parse for maths.
for (const row of items) {
  const chunks = row.transcript.map((l) => ({
    at: l.startTime,
    seconds: Number(l.dur),
    text: l.text.replace(/\[[^\]]*\]|>>/g, '').trim(),
  }));
  console.log(row.id, row.selected.languageCode, chunks.length, 'lines');
}

curl -X POST \
  "https://api.apify.com/v2/acts/apidojo~youtube-transcript-scraper/run-sync-get-dataset-items?token=YOUR_APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"startUrls":["https://www.youtube.com/watch?v=_AbFXuGDRTs"],"lang":"en"}'

JSON is the sensible export, CSV will flatten the transcript array into something you cannot read back. customMapFunction is the tidy place to join the lines into one field if that is all you need.

YouTube Transcript Scraper FAQ

Where YouTube Transcript Scraper fits