A Task-Aware Dual-Level Self-Supervised Learning Method for Effective Sound Event Detection
作者
Jun Liu,Qing Gu,Peng-Fei Cai,N. P. Jiang,Yan Song
出处
期刊: [Institute of Electrical and Electronics Engineers] 日期:2026-04-21卷期号:: 14782-14786
标识
DOI:10.1109/icassp55912.2026.11462317
摘要
The semi-supervised learning for sound event detection (SED), with audio tagging (AT) as auxiliary task has proven highly effective, while self-supervised learning (SSL) also shows strong potential. Thus, exploring SSL methods that complement the semi-supervised learning is promising, yet most existing SSL methods employ single-level pretext tasks (e.g., clip- or frame-level), mismatched with the joint SED-AT paradigm. To tackle this, we propose a task-aware dual-level self-supervised learning method to match the joint SED-AT semi-supervised learning. Specifically, a Transformer-based Siamese network is designed to effectively transfer knowledge to the downstream task by jointly leveraging SED-specific frame-level and complementary AT-specific clip-level objectives via a self-distillation way, thereby enabling task-aware learning. Furthermore, for the frame-level objective, a codebook of prototypes is learned online to provide more effective pseudo-supervision tailored for SED. Experiments conducted on the DESED dataset demonstrate the superior performance of our method, achieving 0.611/0.819 PSDS1/PSDS2 respectively.