Low-complexity bone conduction-aided speech enhancement leveraging sub-band and full-band architecture

This paper presents a low-complexity speech enhancement model that jointly utilizes sub-band and full-band features of bone-conducted (BC) and air-conducted (AC) speech. A two-step speech enhancement procedure is considered. In the first step, the proposed model estimates a set of gains of pre-defined sub-bands to enhance the spectral envelope of noisy AC speech, whereas in the second step, the proposed model estimates a set of coefficients to apply deep filtering to the output of the previous step. Specifically, sub-band gains and deep filtering coefficients are estimated from sub-band and full-band features of both AC and BC speech by two sets of encoder-decoder. Additionally, attention-based feature fusion modules are employed to fuse features of AC and BC at an early and intermediate stage. The multiply-accumulate operations per second of the proposed model are 40.3 M, which is lower than that of most existing BC-aided speech enhancement methods. Experimental results on AC and BC speech recorded from 109 individuals demonstrate that the proposed model effectively balances the noise reduction performance and the computational complexity.