Integrating VGGSK and BEATs for Enhanced Sound Event Detection: A Semi-Supervised GRU-Based System with Weak Labels and Synthetic Soundscapes

This paper presents the architecture we developed for the Detection and Classification of Acoustic Scenes and Events (DCASE) 2023 Challenge, specifically Task 4 on Sound Event Detection using Weak Labels and Synthetic Soundscapes. We integrated embeddings from VGGSK and BEATs, and employed a GRU-based model to classify sound events in each time frame. The system employs thresholding and smoothing techniques in its post-processing phase. For semi-supervised learning, we used the mean teacher approach with an Exponential Moving Average (EMA) strategy to update the teacher model’s parameters. Pseudo-labels, generated by the student model, help leverage unlabeled data. Additionally, we applied data augmentation methods including mix-up, Gaussian noise, and embedding masking. With additional training data, our system achieved a Polyphonic Sound Detection Score (PSDS) of 0.529 for PSDS1 and 0.78 for PSDS2 on the validation dataset.