How to get YouTube video transcripts from a keyword search

The job: you want to know what the popular videos on a topic actually say, not just their titles. Search YouTube, keep the videos worth reading, and pull their transcripts. You end with one JSON Lines file: per video, its title, channel, views, the language of the track you got, the full text for a model to read, and timed segments for quoting with a timestamp.

No Google API key, OAuth consent or proxy is involved. Each call is an HTTPS request with your key.

Step by step

Step 1: find the videos

Call YouTube video search with query. sort takes relevance (the default), latest or most_viewed. One call returns about 20 videos for $0.004, each with the 11-character id, title, url, published_at, the channel in author and metrics.views. Add view=full and each video carries ext.length_s, its length in seconds, which helps you skip Shorts when you want long explanations.

curl -G "https://api.okdata.online/v1/youtube/videos/search" \
  -H "Authorization: Bearer $OKDATA_API_KEY" \
  --data-urlencode 'query=sourdough starter' \
  -d sort=most_viewed -d view=full

Search results leave text empty and metrics.likes and metrics.comments null; YouTube video detail has them if you need them. published_at is worked out from labels such as "3 weeks ago", so treat it as approximate.

Step 2: fetch each transcript

Pass a video's id to YouTube video transcript. One call returns one transcript for $0.004.

• lang is your preferred language, such as en, es or ja. It is a preference: the transcript's own lang field says which track came back, and it is und when the track's language cannot be told. Check it before you file the text under a language.
• segments is the list of timed pieces, each with start and end in seconds and its text. Use it for quotes, chapter notes, subtitles or links that jump to a moment (&t= plus the start in whole seconds).
• text is all segments joined into plain text. Hand this to a model when you want a summary.

curl -G "https://api.okdata.online/v1/youtube/videos/transcript/get" \
  -H "Authorization: Bearer $OKDATA_API_KEY" \
  -d id=Qm7rT2vXa9c -d lang=en

A video without captions returns not_found, which is free, so the bill is lower for every captionless video in your batch.

Step 3: put it together in Python

import json
import os
import requests

API = "https://api.okdata.online"
HEADERS = {"Authorization": f"Bearer {os.environ['OKDATA_API_KEY']}"}
PICK = 5  # transcripts to fetch; each is one call


def get(path, **params):
    return requests.get(API + path, headers=HEADERS, params=params, timeout=90).json()


found = get("/v1/youtube/videos/search", query="sourdough starter", sort="most_viewed", view="full")
if found["code"] != "ok":
    raise RuntimeError(f"{found['code']}: {found['message']}")
videos = sorted(found["data"]["items"], key=lambda v: v["ext"]["length_s"] or 0, reverse=True)[:PICK]

with open("transcripts.jsonl", "w", encoding="utf-8") as out:
    for v in videos:
        body = get("/v1/youtube/videos/transcript/get", id=v["id"], lang="en")
        if body["code"] == "not_found":
            print(v["id"], "no captions, not charged")
            continue
        if body["code"] != "ok":
            print(v["id"], body["code"], body["message"])
            continue
        t = body["data"]["item"]
        out.write(json.dumps({
            "id": v["id"], "title": v["title"], "channel": v["author"]["name"], "views": v["metrics"]["views"],
            "lang": t["lang"], "text": t["text"], "segments": t["segments"],
        }, ensure_ascii=False) + "\n")

A failed transcript is printed and skipped rather than stopping the run, because the other videos are still worth having.

Good to know

• No captions, no transcript. The endpoint reads caption tracks; it does not run speech-to-text. For a captionless video, use Whisper or another speech model on audio you are allowed to use. New uploads sometimes get automatic captions a day later.
• One language per call. Asking for a second language is a second charged call. Ask once, check lang, and translate segment by segment yourself if needed; the timings stay valid.
• Automatic or typed captions. The response does not say which kind of track it is. Treat automatic captions with care when you quote.
• The official API. YouTube's captions.download costs 200 quota units per call and requires permission to edit the video, so it serves your own channel, not other people's videos.
• Retries without paying twice. Send an Idempotency-Key header, such as transcript- plus the video ID. A repeated request with the same key within 24 hours returns the stored answer with no new charge.

If a library such as youtube-transcript-api tells you transcripts are disabled, see what TranscriptsDisabled means. For hundreds of videos at once, see bulk YouTube transcripts.

{
  "endpoint": "youtube.videos.search",
  "params": {
    "query": "claude code tutorial"
  },
  "idempotency_key": "<a new UUID>"
}