Extract Hardcoded Subtitles (OCR)

Read burned-in subtitles from a video into SRT, VTT, ASS and more, in your browser

Runs in your browser

Drop a video with burned-in subtitles

MP4, WebM, MOV or any video your browser can play

The video stays on your device. Only the OCR models (about 25 MB, from Hugging Face) are downloaded once and cached.

What is the hardcoded subtitle extractor?

Hardcoded (burned-in) subtitles are part of the video picture, so they can't be switched off or copied like a subtitle track. This tool plays through the video frame by frame in your browser, finds the text in the area you mark with PaddleOCR text detection, reads it with a PP-OCRv6 recognition model (Tesseract for Vietnamese), and joins frames that show the same text into timed lines. Lines read with low confidence are marked so you can correct them before saving as SRT, VTT, ASS, TTML, LRC or plain text.

Common uses

  • Get an editable subtitle file from a video that only has burned-in captions
  • Translate a clip: extract the original lines, then translate the SRT
  • Copy quotes or lyrics shown on screen as text
  • Re-time or restyle old subtitles in a subtitle editor

When to use it

Use it when the subtitles are drawn on the picture. If the video file has a separate subtitle track, extract that track instead, it is exact. If there are no subtitles at all, use Audio and Video to Subtitles to transcribe the speech.

Examples

  1. English clip to SRT

    Input

    A 3-minute MP4 with white subtitles at the bottom. Language: English, Chinese, Japanese, Latin scripts. 4 frames per second.

    Output

    About 40 timed lines in an .srt file; glued words such as Ithought are split and Al is corrected to AI. A few lines are marked Unsure to check.

  2. Vietnamese subtitles

    Input

    A drama episode with yellow Vietnamese subtitles. Language: Vietnamese.

    Output

    Vietnamese lines with tone marks, ready to download as SRT or VTT.

Questions and answers

Is my video uploaded?

No. Frames are decoded and read in your browser. Only the OCR models are downloaded, from Hugging Face and jsDelivr, and then kept in the browser cache.

Why are some lines marked Unsure?

The recognizer reports how confident it is for every character. Lines with a low average or a single doubtful character are marked so you can compare them with the video: click the time to jump there, then edit the text.

How long does it take?

Roughly a third to the full length of the video at 4 frames per second, depending on your computer. Frames that show the same text as the previous one skip recognition, and WebGPU makes it faster where available.

How can I get better results?

Fit the blue box tightly around the subtitles so logos and signs stay outside, pick the right language, and use 8 frames per second for fast dialogue. Very small, blurry or stylised fonts are read less reliably.

Sources and standards

All Tools