Low-SNR speech enhancement in driving environment
Yang Hongxiao, Jie Wei, Xiaofeng Zhong · 2016
Driving environment is a complex acoustic environment with a variety of noises. It requires more precise speech enhancement of noisy signals, for speech recognition when driving. In this paper, a new two-stage neural network speech enhancement algorithm is proposed. First, the feature vector of noisy signals is used to training BP network and RBP network. And then, the BP and RBP networks are merged for denoising model. In order to improve the robustness of the algorithm, the paper also introduces voice activity detection algorithm based on Energy Statistical Complexity at enhanced stage, which judged the intensity of noise through the algorithm to classify denoising. Experimental results show that the proposed algorithm has better performance on both SNR and intelligibility metrics.