Thank you for PaddleOCR - Built a real-time hardsub OCR browser extension for CJK video immersion

[jmcoder1]

Big thanks to the PaddleOCR team for the incredible work on models and datasets for CJK character detection and recognition!

I built YuzuLingo, a browser extension designed to help language learners read hardcoded subtitles (hardsubs) on web video platforms like YouTube, Bilibili, and streaming sites.

A major challenge with Asian drama and news streams is that captions are burned directly into the video frames. Because there’s no DOM text layer, standard hover dictionaries fail completely.

To solve this, I leveraged PaddleOCR detection/recognition pipeline models exported to ONNX format to run inference directly in the browser via ONNX Runtime Web / WebAssembly. When a user hovers over a subtitle line on a <video> element, it crops the canvas region and runs local OCR to return character bounding boxes and text overlays.

PaddleOCR’s high accuracy on dense Traditional Chinese ($正體字$) and CJK characters made real-time subtitle extraction possible.

Thank you again for open-sourcing these models!

interact_multi_subs_ja.mp4