A low-computational DNN-based speech enhancement for hearing aids based on element selection

In this study, we propose a low-computational deep neural network (DNN)-based speech enhancement scheme for hearing aids. Since the computational resources in the digital signal processor embedded in a hearing aid are very limited, we reduce the input feature dimension for the DNN. To achieve low computational processing, we consider a dimensionality reduction by element selection. The elements are selected by minimizing the reconstruction error of a linear autoencoder. Because it is a selection, the proposed dimensionality reduction does not need any multiplications, unlike other dimensionality reduction algorithms such as principal component analysis. Therefore, our algorithm can reduce computational cost with little degradation of the speech enhancement performance. We evaluate the performance and the computational cost of the proposed algorithm compared with conventional algorithms.