PixelOkay
● Local video processing
Tools

PIXELOKAY / VIDEO & AI TOOLS

Video Text Extractor

Find text in video frames. Capture timestamps, merge repeated titles, search and correct results, then export subtitles or a structured report—without uploading footage.

Local OCRTimestamped textSearchable timelineSingle-frame rescueBatch correctionsSRT · VTT · CSV · JSON
Upload a video or try the example to begin. OCR loads only when you choose Scan.

Video preview

Scrub, crop a text area or jump to a detected occurrence.

No source
Your footage appears here.Load a video to start scanning.
0.0 / 0.0 s

Visible text OCR is not speech transcription. SRT/VTT exports mark when sampled screen text appears, not when someone speaks.

04

Text timeline

0Segments
0Observations
0Low confidence
Filtered results

Batch actions affect all matching results after search and filtering, even if only the first 50 are currently rendered. Undo restores the previous state. Invalid timing shifts are rejected.

No text segments to review.

No recognized text yet.

Run Scan, or import an existing PixelOkay OCR report to review and export it.

AI

Optional AI text correction review

Tesseract reads video frames; this separate, local language model only suggests typo corrections to existing OCR text. It cannot verify pixels or timestamps. Review every suggestion before applying.

Nothing is sent to a server for inference. AI model files are downloaded from third-party CDNs only after you approve setup. Suggestions are not OCR evidence and will never be applied automatically.

05

Export & projects

Import a JSON report with a cues array, or SRT/WebVTT subtitles for correction. Imported text can be reviewed without OCR downloads. Video files are not included in reports. Timing remains editable.

How to extract text from a video

1

Load and focus

Choose a local video and draw a selection around the area you want read. Subtitles and slide titles benefit from a focused crop.

2

Scan and review

Choose frame interval and OCR language. The scanner merges repeated sightings; use single-frame recognition to recover a missed title.

3

Edit and export

Correct text and timing, batch include/exclude or shift filtered results, and download SRT, WebVTT, text, CSV or JSON.

Video OCR vs. speech subtitles

OCR reads words visible in frames—slides, signs, labels, graphics and on-screen captions. It does not transcribe spoken audio. Because the video is sampled rather than analyzed every frame, brief flashes of text between samples may not be found. For frame-accurate captions, inspect and adjust the generated times manually.

Using Video Text Extractor

Extract visible on-screen words from sampled video frames, review their approximate timing and export edited text cues. OCR reads pixels and does not transcribe spoken dialogue.

  1. Load a video or demo, choose the language, crop and bounded time range that contain the text.
  2. Run OCR, inspect the sampled cues and use the single-frame rescue control for a missed title card.
  3. Correct text and cue timing, choose which cues to include, then export SRT, VTT, TXT, CSV or editable JSON.

Example: For a clip with a lower-third name card, crop to the title area, scan the short range where it appears, correct the recognized name and cue start, then export an SRT sidecar.

Limits and privacy: A scan is limited to 180 seconds and 240 sampled frames, so brief overlays can be missed and cue timing is approximate. The first scan retrieves Tesseract.js worker, WASM and language assets from external hosts; optional text correction also downloads model files after explicit consent. These requests may disclose the visitor IP. OCR reads visible pixels rather than speech, and advanced subtitle styling may not round-trip; project JSON preserves excluded cues. One bounded single-frame English OCR check succeeded; full-range scan timing, grouping, other-language coverage and broader accuracy remain unverified.

Frequently asked questions

Does this transcribe speech in the video?

No. OCR reads words visible in sampled frames. Spoken dialogue requires a separate speech-recognition workflow.

Why are some titles missing or timed approximately?

The tool samples up to 240 frames in a bounded range rather than reading every encoded frame. Use single-frame rescue for a missed title and adjust cue timing during review.

Related tools

LOCAL AI SETUP

Set up AI on this device

First, a model needs to be loaded on your device

Optional OCR text corrections run locally in your browser after you explicitly load a model. The model may require hundreds of megabytes of CDN files. Compatible cached files can be shared with other PixelOkay tools on this same origin. Nothing downloads automatically on entry.

  1. 1. Check device
  2. 2. Download / cache
  3. 3. Initialize model
  4. 4. Review OCR fixes
Choose a compatible WebGPU model for OCR text corrections.

Device & model resources

Approximate measurements
Device RAM / model memoryChecking device information…

Live system RAM, GPU memory and actual model working-memory requirements are unavailable in most browsers. Model download size is not the RAM requirement.

Checking device compatibility…

Browser website storageChecking storage…

Site-wide storage estimate, includes other PixelOkay tools. Not RAM.

Model cache not yet inspected.

ONE ACTION · ALL RECOGNIZED AI MODEL DOWNLOADS

Free AI model storage across PixelOkay

Scan this website origin for WebLLM and Transformers.js model caches, plus recognizable model files in other cache buckets. Includes other PixelOkay tools on this origin. Unrelated site data is not cleared automatically.

Run Inspect all AI caches to review safe model-only targets.

Unknown caches, browser-native models, inaccessible IndexedDB and OPFS files are not deleted; close other active AI tabs before cleanup.

Having trouble downloading or loading AI?

First try Refresh Storage & Cache, then retry or select a smaller compatible model. If necessary, use the visible cache controls immediately above to inspect and remove verified model-only downloads. Other PixelOkay tools may need to redownload those models. Do not clear all website data without backups.

Select a model, then check compatibility. The regular OCR scanner always works without AI.
Currently loadedNoneNo active inference session
Technical details

WebLLM 0.2.87 · WebGPU. Runtime and model files may load from CDNs. Inference on OCR text is on-device. No API keys or server-side inference.

Remove recognized model caches?

Review the recognized model-cache targets before deletion.

Only identified model-cache entries are targeted. Saved logs, chats, projects, editor work, and unrelated site caches are not targeted; the scanner cannot inspect all browser storage providers.