USAD 2.0: Scaling Representation Distillation for Universal Audio Understanding

arXiv·medium signal

USAD 2.0 is a universal audio encoder that distills knowledge from both self-supervised and supervised teachers to cover speech, music, and general audio in a single encoder — important as audio LLMs increasingly rely on one encoder for diverse inputs. It improves on prior multi-domain encoders like USAD and SPEAR in both coverage and evaluation breadth. Relevant for builders working on audio-language model front-ends.