Joint analysis of acoustic scenes and sound events based on multitask learning with dynamic weight adaptation

Acoustic scene classification (ASC) and sound event detection (SED) are major topics in environmental analysis. Considering that acoustic scenes events closely related to each other, the joint analysis of using multitask learning (MTL)-based neural networks was proposed some previous works. Conventional methods train MTL-based models a linear combination ASC SED loss functions with constant weights. However, performance conventional depends strongly on weights losses, it is difficult determine appropriate balance between losses MTL SED. In this paper, we thus propose dynamic weight adaptation for based average (DWA) multi-focal (MFL) adjust automatically. By comparing two methods, then clarify how weights, rather than specific DWA MFL, generally benefits MTL. Moreover, investigate training model dynamically progresses disclose affect their performance.