Reproduce “A Hybrid System of Sound Event Detection Transformer and Frame-wise Model for DCASE 2022 Task 4” on OpenResearch

About this paper

Yiming Li, Zhifang Guo, Zhirong Ye, Xiangdong Wang, Hong Liu, Yueliang Qian, Rui Tao, Long Yan and 1 more

This paper presents a hybrid system for weakly supervised Sound Event Detection, integrating an event-wise Transformer and a frame-wise CNN model. The system achieved competitive performance in DCASE 2022 Task 4, demonstrating that combining these distinct architectures with advanced semi-supervised learning techniques and class-specific fusion strategies improves detection accuracy and boundary estimation.

Key insights

• Event-wise (SEDT) and frame-wise (MLFL-CNN) models offer complementary strengths for SED, with one excelling in metrics sensitive to false positives and the other in metrics robust to temporal errors.
• Sophisticated semi-supervised learning techniques, including online pseudo-labeling with EMA, asymmetric augmentation, Mixup, and Focal Loss, are crucial for effectively leveraging heterogeneous and unlabeled data.
• Class-specific model fusion and post-processing with optimal window tuning can significantly boost overall system performance by tailoring optimizations to individual sound event categories.

Summary by alphaXiv.