Acoustic Scene Classification for Bone-Conducted Sound Using Transfer Learning and Feature Fusion

Abstract

The air-conducted (AC) sound is usually used in the task of acoustic scene classification (ASC). Compared with the AC sound, bone-conducted (BC) sound has the unique advantage of shielding background noise. However, the amount of information contained in BC sound is far less than that in the AC sound due to its limited frequency bandwidth. In this paper, an acoustic scene classification method for BC sound is proposed with a small BC dataset. Firstly, the prosodic features are combined with the spectral features to capture more information, and feature fusion is adopted. Secondly, in order to deal with the small BC dataset, transfer learning is used with a large AC dataset. Finally, a deep learning network based on local residual learning is proposed. The experimental results show that the proposed method achieves the superior performance over the reference models.