I've been trying different OCR models on what should be very simple - subtitles (these are simple machine-rendered text). While all models do very well (95+% accuracy), I haven't seen a model not occasionally make very obvious mistakes. Maybe it will take a different approach to get the last 1%...