Conditional Mixup SED model achieves 0.645 PSDS1 on DESED
New semi-supervised method beats previous SOTA by 3% on DESED validation set.
Sound event detection (SED) is critical for environmental analysis but suffers from scarce labeled data. While large pretrained audio foundation models exist, fine-tuning them is challenging when labeled data is limited. Previous work ATST-SED used a pseudo-label based semi-supervised fine-tuning framework. In this paper, the authors improve on ATST-SED by introducing an embedding-level self-supervised contrastive loss inspired by ATST-Frame pretraining. This loss better exploits unlabeled data during fine-tuning.
A key challenge: mixup serves different roles in the two objectives—composition mixup for pseudo-label learning and perturbation mixup for contrastive learning. To resolve this mismatch, the authors propose conditional mixup, which combines both forms within a single semi-supervised framework. They also define corresponding embedding-level contrastive losses. The resulting model achieves 0.645 PSDS1 and 0.822 PSDS2 on the DESED validation set, establishing a new state of the art. The paper was accepted at SMC 2026.
- Achieves 0.645 PSDS1 and 0.822 PSDS2 on the DESED validation set, new SOTA.
- Introduces conditional mixup to unify composition and perturbation mixup in one framework.
- Leverages embedding-level contrastive loss from ATST-Frame pretraining to better use unlabeled data.
Why It Matters
Enables more accurate environmental sound detection with less labeled data—critical for smart cities, surveillance, and acoustic monitoring.