Exploiting lightweight neural post-filtering for directional speech enhancement

Abstract

Directional speech enhancement using a microphone array aims to isolate the target speaker's voice from competing speakers and background noise. While traditional signal processing-based beamformers often struggle to suppress residual interference, end-to-end methods face challenges in operating on low-resource devices and generalizing to real-world scenarios. To address these issues, we propose a hybrid approach that integrates the robust generalized sidelobe canceler (GSC) with a neural post-filter. Our method employs a lightweight network architecture as the backbone of the neural post-filter, utilizing both microphone array signals and GSC outputs, which are modeled by separate encoders. Additionally, we introduce an auxiliary decoder to learn the target components in the GSC outputs, thereby enhancing the post-filter's performance when combined with a knowledge distillation training strategy, all without introducing additional computational load during inference. Experimental results on simulated and real-world datasets demonstrate the superior performance of the hybrid system, validating the effectiveness of our proposed methods.