Computational audition aims to enable artificial systems to process, analyze, and understand the acoustic world in a manner analogous to human listening. While deep learning has driven remarkable breakthroughs in acoustic analysis, standard supervised learning paradigm depends heavily on massive collections of pristine, precisely time-stamped, and verified ground-truth labels — annotations that are prohibitively expensive, subjective, and scarce in real-world acoustic settings.

This dissertation investigates and develops foundational machine learning frameworks for learning from imprecise acoustic labels, including weakly supervised labels (audio-level tags without onset/offset timestamps), webly scraped annotations, noisy/corrupted labels, partial labels, and complementary supervision. We establish theoretical bounds, design robust loss-weighting and sampling strategies, and validate our methods across large-scale benchmarks (Audioset, DCASE challenges, public safety gunshot diagnostics, and real-time smart surveillance).