Single Channel Speech Enhancement Algorithm Based on BLSTM-DNN Bidirectional Optimized Hybrid Model

Xiaoyue Sun,Ruwei Li,Tao Li,Dengcai Yang
DOI: https://doi.org/10.1088/1757-899x/719/1/012027
2020-01-01
Abstract:The performance of existing speech enhancement algorithms based on deep learning is not ideal in complex noise environment. To improve the problem, a bidirectional optimized hybrid network named BLSTM-DNN is constructed based on bidirectional long-short term memory (BLSTM) network and fully-connected deep neural network (DNN). This structure uses BLSTM to extract high-level information including past and future temporal context of noisy speech. Next, fully-connected DNN fits the high-level information to ideal ratio mask (IRM). Finally, the IRMs estimated by the BLSTM-DNN are used to enhance the noisy speech. Experimental results show that the proposed method can effectively improve the speech quality and intelligibility under unknown noise conditions.
What problem does this paper attempt to address?