Disentangled Representation Learning for Tackling the Long Tail Problem How Secure are Deep Optical Privacy Preserving Methods? Active Learning and Task-Adaptive Pre-Training for Low-Resource Language Transcription