Video OCR - Extract On-Screen Text from Video (Captions)

What does Video OCR do?

Video OCR reads the on-screen text of a video file so you can extract text from video without watching it: burned-in subtitles, captions, ad hook text, titles, lower-thirds (names and job titles), slide and screen-recording text, prices and product names. You give it direct video links (or the frames of a Video Frame Extractor run) and get back every distinct piece of text once, with the time it appears and disappears, where it sits on the screen, a frame image, an LLM-ready transcript and an SRT/VTT on-screen text track.

A caption shown for 4 seconds is one row, not four: the same text in the same place over consecutive frames is merged into one text segment, even when OCR misreads a letter in one frame or the caption builds up word by word.

Why use this video OCR API?

• Video-native, not image OCR. One run decodes the video, samples a frame every second (your choice), skips frames that did not change, reads the text and merges it over time. Image OCR actors make you do all of that yourself.
• You only pay for frames that contain text. Frames with no text, blank frames, unchanged frames, failed and refused videos are free. A video without any on-screen text costs nothing.
• Timestamps you can trust: startSeconds / endSeconds per segment, accurate to the sampling interval (1 s by default).
• Ready for AI and automation: fullText (one line per segment with its time), hookText (text shown in the first 3 seconds), SRT and VTT files, clean JSON for n8n, Make, Zapier or your own code.
• No API keys, no data sent to third parties: open-source PP-OCR models (Apache-2.0) run inside the Actor.
• Chains with Video Frame Extractor: pick a finished Video Frame Extractor dataset and the text of those frames is read with their original timestamps.

Use cases

• Burned-in subtitles to text / SRT: recover subtitles that only exist inside the picture (hardsubs), e.g. to translate or index them.
• Ad and UGC teardown: pull the hook text, claims, prices and call-to-action overlays of ad creatives and short-form videos you have the files for.
• Lower-thirds and names in news, interviews and webinars: who appears when, with their title.
• Slides and screen recordings: searchable text of every screen, with time ranges.
• Product text in videos: brand names, packaging text, prices, phone numbers and URLs shown on screen.
• Video search and RAG: put the on-screen text next to the transcript so an LLM sees what the viewer reads.

How to extract text from a video

1. Paste one or more direct video file links into Video URLs (MP4, MOV, WEBM, MKV, AVI, M4V, MPEG-TS; your own storage, S3/R2/GCS, an Apify key-value store, Dropbox or Google Drive "anyone with the link" share links).
2. Choose how often a frame is read (every 1 s is right for subtitles and captions; 0.5 s for fast word-by-word captions; 2-5 s for slides and long screen recordings).
3. Pick the text language (script) and, optionally, a text area (bottom = subtitles and lower-thirds only).
4. Click Start. Text segments appear in the On-screen text view; the Videos view has the transcript and SRT/VTT links.

Pages of video platforms (YouTube, TikTok, Instagram, Facebook, X, Vimeo, Twitch...) are refused politely, free of charge: their terms forbid downloading. Use files you are allowed to process.

Input example

{
  "videoUrls": ["https://example.com/ads/summer-sale.mp4"],
  "sampleIntervalSeconds": 1,
  "ocrLanguage": "en",
  "textRegion": "full"
}

Chain from Video Frame Extractor: leave Video URLs empty and pick its run's dataset in Video Frame Extractor dataset, or paste its frame links into Frame image URLs.

Output example

One row per text segment:

{
  "type": "text",
  "fileName": "sample-ocr-captions.mp4",
  "segmentIndex": 5,
  "text": "There are 3 types of
spectra that we can use",
  "startSeconds": 10,
  "endSeconds": 13,
  "durationSeconds": 3,
  "timecode": "00:00:10.000",
  "endTimecode": "00:00:13.000",
  "timestampSeconds": 10,
  "frameUrl": "https://api.apify.com/v2/key-value-stores/.../records/ca66bc8e-sample-ocr-captions-0007-000010000ms.jpg?signature=...",
  "confidence": 0.9906,
  "box": { "x": 0.1444, "y": 0.6695, "width": 0.7097, "height": 0.0812 },
  "position": "bottom",
  "lineCount": 2,
  "framesSeen": 2,
  "isPersistent": false
}

fullText of the same video (transcript style, one line per segment):

[0:01] If you do a quick search for spectroscopy data,
[0:04] there are numerous ways that the data can appear.
[0:07] Those variations are based on the source of the cosmic light.
[0:10] There are 3 types of spectra that we can use
[0:13] Continuous Spectrum

Plus one summary row per video (type: "video") with status (ok, no-text, failed, skipped), segmentCount, framesWithText (the frames charged), hookText, fullText, srtUrl, vttUrl, and frames[] with index, timestampSeconds, timecode, url and text of every frame that had text - the same field names as Video Frame Extractor, so the two join on timestampSeconds.

box is in fractions of the frame (0-1). isPersistent marks text that stays on screen for most of the video (watermarks, channel logos, "Artist's concept" labels) so you can filter it out.

How much does it cost to extract text from a video?

Pay per event: $0.002 per frame with text ($2 per 1,000 frames), plus Apify's tiny actor-start fee. Store discounts: Silver $0.0018, Gold $0.0016 per frame.

Free: frames where no text was found, blank frames, frames identical to the previous one, failed or refused videos. A video with no on-screen text at all costs $0.

Worked examples, measured on real videos (1 frame per second unless noted):

Video	Frames read	Frames with text	Cost
30-second vertical explainer, captions most of the time (the built-in sample)	29	24	$0.048
37-second promo with burned-in English subtitles	32	32	$0.064
71-second vertical promo: big title text + name lower-thirds	70	44	$0.088
33-second TV ad, text only on a magazine cover and the end card	30	7	$0.014
60-second interview, no on-screen text	60	0	$0.00
3.6-minute screen recording, one frame every 2 s (unchanged screens skipped)	26	26	$0.052
18.7-minute subtitled documentary (600 frames spread over the video)	387	382	$0.76

Max frames per video (default 600) caps the cost of one video: longer videos are sampled evenly instead. Max cost per run is respected exactly: reading stops when the budget is reached, the video is marked truncated, and later videos get a free skipped row.

Languages

en reads English and also Chinese and Japanese. Other options: Latin with accents (French, German, Spanish, Portuguese, Italian, Czech, Polish...), Cyrillic (Russian, Ukrainian, Bulgarian...), Korean, Arabic, Devanagari (Hindi, Marathi...), Greek, Thai.

Tips for best results

• Subtitles only: use Text area = bottom - signs, logos and UI text elsewhere in the picture are ignored and the run is faster. For small or thin subtitles also try Minimum confidence 0.6.
• Word-by-word TikTok captions: set Read a frame every to 0.5 s; the growing caption is merged into one segment with its full sentence.
• Slides and screen recordings: 2-5 s is enough and much cheaper; unchanged screens are skipped for free anyway.
• Stylised fonts can produce letter errors (for example 0 vs O). Check confidence and framesSeen; a text seen in many frames is reliable.

Limits

• Text accuracy is that of a fast CPU OCR model: clean captions, titles and lower-thirds read very well; tiny UI text, handwriting, heavily stylised or moving 3D text can be misread or missed.
• Timing is accurate to the sampling interval (a caption shown for less than one interval can be missed).
• Text that moves (scrolling credits, tickers, tiles sliding across the screen) or sits in a busy moving scene (street signs) produces several short segments instead of one.
• Videos up to 120 minutes / 2 GB by default (raise the limits up to 600 minutes / 4 GB).

Related Kanto Labs actors

• Video Frame Extractor - keyframes, scenes, thumbnails and evenly spaced frames from any video file. Chain its dataset into Video OCR.
• Video to Slides - unique slides of a lecture or webinar as images + PDF with slide text.
• Video to LLM Context - keyframes per scene + transcript, packaged for an LLM.
• Video Scene Detection & Shot List - every cut with timecodes, a shot list and CSV/EDL exports, to pair on-screen text with each shot.
• Video Ad Creative Breakdown - on-screen text plus speech, shots, pacing and CTA in one breakdown per ad or short video.
• AI Image Caption & Alt Text Generator - captions, alt text and tags for images and video frames, to describe what is shown next to the text that is read.

FAQ

Is this the same as transcribing a video?

No. Transcription turns speech into text. Video OCR reads the text shown on screen. Use both for a complete picture: the transcriber for what is said, Video OCR for what is written.

Can I use YouTube or TikTok links?

No. Those platforms forbid downloading in their terms. Use the original files (your own uploads, ad library exports you are allowed to use, your storage).

Does it work from n8n, Make or Zapier?

Yes. Run the Actor (or a saved task) through the Apify integration and read the dataset; the fullText and srtUrl fields are ready to send to an LLM, a sheet or a subtitle tool.

What happens when a video has no text?

It gets a no-text summary row and is not charged.

Something went wrong?

Every problem (404, a web page instead of a file, too large, unreadable file) gets a clear error in its row and is free. If a result looks wrong, open an Issue with the video link and we will look at it.