Speech Enhancement Based on Multi-Objective Ensemble Learning

Abstract

The performance of traditional speech enhancement methods based on deep neural network is limited by using single training objective and network structure. In this paper, we propose a speech enhancement method based on multi-objective ensemble learning. First, the traditional multi-objective learning network structure is modified to reduce the training conflict caused by excess shared parameters. Then, a multi-objective ensemble learning based speech enhancement method is established by employing the modified multi-objective deep neural network (DNN), convolutional neural network (CNN) and gate recurrent unit (GRU), which overcomes the limitation of homogeneity in base models in the traditional ensemble learning based speech enhancement network. The experimental results show that the proposed methods outperforms the traditional multi-objective learning or ensemble learning based speech enhancement methods at the scores of perceptual evaluation of speech quality (PESQ) and short-time objective intelligibility (STOI).