Latent Mixup Knowledge Distillation for Single Channel Speech Enhancement

Abstract

Traditional speech enhancement methods often rely on complex signal processing algorithms, which may not be efficient for real-time applications or may suffer from limitations in handling various types of noise. Deploying complex Deep Neural Network (DNN) models in resource-constrained environments can be challenging due to their high computational requirements. In this paper, we propose a Knowledge Distillation (KD) method for speech enhancement leveraging the information stored in the intermediate latent features of a very complex DNN (teacher) model to train a smaller, more efficient (student) model. Experimental results on a two benchmark speech enhancement datasets demonstrate the effectiveness of the proposed KD method for speech enhancement. The student model trained with knowledge distillation outperforms SOTA speech enhancement methods and achieves comparable performance to the teacher model. Furthermore, our method achieves significant reductions in computational complexity, making it suitable for deployment in resource-constrained environments such as embedded systems and mobile devices.