Light-Weight Causal Speech Enhancement with Bone-Conduction

Abstract

Speech enhancement aims to improve the quality of speech degraded by various types of noise, particularly under challenging conditions such as extremely low signal- to-noise ratio (SNR). Traditional methods predominantly rely on speech data captured by air-conduction (AC), which are highly susceptible to noise. This makes speech en- hancement at low SNRs a challenge. In contrast, bone-conduction (BC) is more robust to noise but provide information constrained to a limited frequency bandwidth. In this paper, we propose a novel fusion module that effectively integrates information from both air-conduction and bone-conduction. Additionally, we introduce a light-weight, causal network designed for low computational complexity, making it suitable for de- ployment on resource-constrained devices. Experimental evaluations demonstrate that the proposed model significantly outperforms the baseline, achieving superior speech quality while reducing model size without an increase in computational complexity.