Acoustic Signal Denoising of Cetacean Based on Treble Band-Split Convolution Network

Ocean ambient noise is known to negatively impact the monitoring of cetacean acoustic signals. Existing deep learning models for generally purpose noise suppression typically process the entire available frequency band of the signals. This is not ideal in the case of cetacean vocalization processing as its spectrum characteristics exhibit almost exclusively in a certain frequency band. Therefore, a hybrid model called Treble Band-Split Convolution Network is designed and proposed in this manuscript specifically for cetacean acoustic signal denoising. This Network nonlinearly divide the entire spectrum into three sections based on the differences in the respective featuring spectrum components of ocean ambient noise and cetacean vocalization, aiming for minimizing the negative impacts from full-band processing protocol. As shown in experimental validations, the proposed model demonstrate significantly improvement on denoising performance compared to traditional method as well as to other approaches based on deep learning network.