I'm a developer learning Mandarin Chinese. I was fed up with trying to watch raw videos on YouTube, Bilibili, and streaming sites, only to realize all the subtitles were burned directly into the video frames.
Because the text wasn't selectable HTML, standard hover dictionaries like Yomitan or Zhongwen completely failed. I had to constantly pause the video, open Pleco on my phone, and draw complex Traditional characters by hand just to look up a single word. It totally killed my watching flow.
So I built YuzuLingo to solve that exact friction.
You just hover your mouse over any hardcoded subtitle or text inside a web video player. The extension runs a local OCR engine directly over the video frame and overlays instant pinyin and dictionary definitions.
For example, if I'm watching a Taiwanese drama or news clip with hardcoded Traditional Chinese captions:
I hover over the character on screen.
The extension OCRs the frame in milliseconds.
I get the definition and reading without pausing or switching apps.
It's free, no account to create, and requires no paid API keys. It works on Firefox, Chrome, Edge, and Brave.
If you're studying English, Spanish, Chinese, Japanese, or Korean through video immersion, try it out and tell me what you'd add or change. I read every comment!
Top comments (2)
Pausing to hand-draw Traditional characters into Pleco mid-scene is such a specific kind of pain, and I can picture how fast it kills the flow of a drama episode. Hovering over the burned-in line and getting pinyin right on the frame is a clever way around it.
Does it keep working in fullscreen? That's where I'd end up using it most, on long episodes where I don't want the browser UI in the way.
This hits a real pain point — hardcoded subs on language content are brutal when you're trying to mine sentences. Curious how you're handling the OCR pipeline: Tesseract with custom traineddata, or something lighter like PaddleOCR? The vertical text layout in Chinese variety shows and dramas used to wreck recognition for me until I added a preprocessing step for text orientation detection. Also, any thought given to exporting the extracted lines with timestamps into Anki format? That'd close the loop for sentence mining workflows PS: the tool I meant is on labagent .tech