Yeah, segmentation is definitely one of the nastiest parts, especially when letters connect fluidly or change shape depending on position—like in cursive Arabic or joined Hindi forms. I’ve noticed CJK can get rough too with all those strokes crammed together under bad mobile lighting. A while back I played around with a few on-device setups for quick scans, and one that stood out handled the script variations surprisingly smoothly without needing cloud calls. If you’re digging for something offline and privacy-focused, https://ocrstudio.ai/ has been pretty solid in my tests for those exact non-Latin cases, though honestly it still takes some tweaking for perfect edge cases. Nothing’s flawless yet, but it beats a lot of the free ones that choke on directionality or tiny marks.