Triad: Tri-Head with Auxiliary Duplicating Permutation Invariant Training for Multi-Task Sound Event Localization and Detection

Sound event localization and detection (SELD) requires simultaneous recognition of events and estimation of their spatial positions. Existing output representations have notable limitations. Single-branch heads couple detection and localization too tightly, so improving one task inevitably affects others. Conversely, conventional multi-branch heads fail to represent overlapping events of the same class, reducing their effectiveness. To overcome these limitations, we propose TriAD, a tri-head output representation for SELD. It combines track-wise localization with Auxiliary Duplicating Permutation Invariant Training (ADPIT), enabling the capture of overlapping events. Experiments on DCASE2025 stereo dataset show that TriAD achieves state-of-the-art F1 score performance. We further conduct the first systematic study of gradient-aware multi-task optimization for SELD and demonstrate that, among all tested strategies, Projected Conflicting Gradient (PCGrad) delivers the largest F1 score gains. These findings highlight the importance of robust representations and effective optimization strategies for advancing SELD systems.