Multi-Task Domain Adaptation with Semi-Supervised Learning for Real-World Acoustic Event Detection

In Acoustic Event Detection (AED) and acoustic denoising tasks, practical application scenarios present substantial challenges, including pronounced discrepancies between training data and real-world data, as well as the difficulty of acquiring large quantities of labeled data. These factors significantly impede the model's generalization performance in real-world environments. To address this issue, we propose SSMT-DAN (Semi-Supervised Multi-Task Domain Adaptation Network), a multi-task domain adaptation model that integrates denoising and classification tasks. SSMT-DAN facilitates end-to-end training through a unified framework, focusing on enhancing model performance in real-world contexts. In Task1, SSMT-DAN employs self-supervised learning to transform noisy audio signals into clean signals and optimizes in-domain feature distribution by introducing an Entropy-Based Rate Reduction (EBRR) strategy, resulting in more compact intra-class samples and more dispersed inter-class samples. Additionally, Pairwise Metric Learning (PML) further accentuates feature distribution differences between domains, thereby enhancing the model's domain adaptation capability. In Task2, SSMT-DAN leverages a small amount of labeled data and a large volume of unlabeled data in the target domain, learning domain-invariant prototypes for each class through semi-supervised Min-Max Entropy Adversarial Training, thereby enhancing the model's classification performance in the target domain. Experimental results demonstrate that SSMT-DAN significantly outperforms existing state-of-the-art methods across multiple datasets, particularly showcasing superior performance and robustness in handling noisy, cross-domain audio data.