YouTube Shorts Text Extractor

The YouTube Shorts Text Extractor is an Apify Actor that reads public YouTube Shorts and returns every piece of visible on-screen text as structured data. Creators and brands burn text, captions, stickers, hashtags, and watermarks directly into their videos — this Actor unlocks that content for search, analysis, and database ingestion without manual review.

This is visual OCR, not audio transcription. Speech-to-text tools give you what was said; this Actor gives you what was shown — the headline card, the price tag, the discount code, the lower-third name badge, the "link in bio" call to action. That text is invisible to every transcript-based tool.

Each Short is processed by AI vision and the results are returned as a structured dataset of text segments — each with its timestamp, type, screen position, language, and confidence rating — plus a combined fullText field for quick search and indexing.

---

Key Features

• On-screen text only: Reads text rendered on top of the video, never the audio track.
• Full coverage: Captures animated overlays, burned-in captions, sticker labels, watermarks, hashtags, and @mentions.
• Timestamp-aware: Records the MM:SS timestamp when each text element first appears.
• Type classification: Distinguishes overlay, caption, sticker, watermark, hashtag, mention, and other text types.
• Combined fullText: All segments joined into a single newline-separated block, ready for full-text search.
• Multilingual: Detects the ISO 639-1 language code for each segment from the text itself.
• Position detection: Classifies text placement as top, center, or bottom of screen.
• Confidence rating: High/medium/low rating based on text clarity in the video frame.
• Flexible URL input: Accepts youtube.com/shorts/..., youtu.be/..., and youtube.com/watch?v=....
• No login required: Works with any public YouTube Short.

---

Why Extract Text from YouTube Shorts?

Shorts are a primary publishing surface for businesses, creators, and brands. Product drops, discount codes, event announcements, hiring notices, and pricing updates are routinely shared only as burned-in video text — never structured, never indexed, and never captured by transcript tools. This Actor makes that content machine-readable for:

• Content intelligence and brand monitoring tracking what competitors publish on Shorts.
• Retail and e-commerce indexing product drops, promo codes, prices, and launch dates.
• Food and hospitality capturing daily specials and seasonal menus announced via video.
• Event aggregators extracting event names, dates, venues, and lineup text from promo Shorts.
• Creator and influencer analytics measuring hook text, captions, and call-to-action patterns.
• Market research tracking pricing, offers, and messaging trends across channels.
• Accessibility tools converting visual Shorts text to readable formats.

---

Who Is It For?

• Marketing and analytics teams monitoring brand or competitor Shorts content at scale.
• Product and data teams building structured datasets from short-form video.
• Developers integrating Shorts text extraction into discovery or monitoring pipelines.
• Researchers studying visual communication trends in short-form video.

---

Input Schema

{
  "shortUrls": [
    "https://www.youtube.com/shorts/0uTo-oe2ecA",
    "https://www.youtube.com/shorts/MJstrGQT79k"
  ],
  "maxItems": 10
}

shortUrls is required. All other fields are optional.

Field	Type	Default	Description
shortUrls	Array of strings	—	Public YouTube Shorts URLs (/shorts/..., youtu.be/..., or /watch?v=...)
maxItems	Integer	10	Maximum number of Shorts to process in a single run
proxyConfiguration	Object	Residential on	Proxy used for downloads — YouTube blocks most datacenter IPs

---

Output Schema

Each dataset item represents one Short and all the on-screen text found in it:

{
  "shortUrl": "https://www.youtube.com/shorts/MJstrGQT79k",
  "videoId": "MJstrGQT79k",
  "duration": 45,
  "textSegments": [
    {
      "text": "Suffolk wildfires ‘stabilising’, say firefighters",
      "timestamp": "00:00",
      "type": "overlay",
      "position": "top",
      "language": "en",
      "confidence": "high"
    },
    {
      "text": "At least 12 fire engines remain on scene.",
      "timestamp": "00:09",
      "type": "caption",
      "position": "bottom",
      "language": "en",
      "confidence": "high"
    }
  ],
  "fullText": "Suffolk wildfires ‘stabilising’, say firefighters\nAt least 12 fire engines remain on scene.",
  "scrapedAt": "2026-08-01T07:16:28.904Z"
}

Field	Description
shortUrl	Normalized canonical URL of the Short
videoId	YouTube 11-character video ID
duration	Video length in seconds
textSegments	Array of all on-screen text elements found
textSegments[].text	The visible text content as it appears on screen
textSegments[].timestamp	MM:SS when the text first appears — null if indeterminate
textSegments[].type	overlay | caption | sticker | watermark | hashtag | mention | other
textSegments[].position	top | center | bottom — vertical screen position, null if it moves
textSegments[].language	ISO 639-1 language code detected from the text, e.g. "en", "es" — null if indeterminate
textSegments[].confidence	high | medium | low — extraction confidence based on text clarity
fullText	Every segment's text joined with newlines — convenient for search and indexing
scrapedAt	ISO timestamp of when this Short was processed

---

How It Works

1. Validate inputs — each URL is checked against accepted YouTube URL patterns, the video ID is extracted, and the URL is normalized to a canonical Shorts form.
2. Fetch (retriable) — each Short is retrieved in a dedicated crawler step. Failed fetches are retried automatically with a fresh proxy (up to 5 attempts).
3. Process — the video is analyzed by AI vision to extract all visible on-screen text in a single pass.
4. Structured output — each text segment is classified by type, position, language, and confidence.
5. Push to dataset — one dataset row is pushed per Short containing all its text segments plus the combined fullText.

---

Pricing

This Actor uses pay-per-event pricing:

Event	Charged
short-processed	Once per Short successfully processed
video-second	Per second of video processed (rounded up, minimum 1)

Shorts that fail to download or contain no on-screen text are not charged.

---

Notes and Limitations

• Public Shorts only: Private, unlisted, and age- or login-gated videos are not supported.
• Video availability: Deleted, region-blocked, or expired videos will fail to fetch and are skipped with a warning.
• Proxy recommended: YouTube blocks most datacenter IPs. Residential proxy is the default and is strongly recommended.
• OCR accuracy: Fast-moving, small-font, or low-contrast text may yield lower confidence extractions.
• Text-free videos: Shorts with no burned-in text return no dataset row — this is expected, not an error.
• Video size: Very large videos are skipped automatically.
• Audio not included: Spoken content is intentionally excluded — only text visually rendered on screen is extracted. Use a transcription Actor if you need speech.