[2Win5] Poster session 2

[2Win5-54] M2D-X: Towards a Universal Audio Pre-training Framework

〇Daisuke Niizumi¹, Daiki Takeuchi¹, Yasunori Ohishi¹, Noboru Harada¹, Kunio Kashino¹ (1.NTT Corporation)

Keywords:audio representation, representation learning

General-purpose audio representations are useful building blocks for various audio applications; however, specialized representations for tasks that learn from application task data can be more useful. This study proposes M2D-X, a general framework for learning application-specific audio representations. Experimental results show that the proposed M2D-X learns effective representations with top-level performance for the highly competitive AudioSet and speech domain, a small-data medical task, and a zero-shot classification task.