摘要
针对噪声的随机性和突变性,使得传统算法抑制非平稳噪声比抑制平稳噪声难度增大的问题,提出了一种基于深度神经网络的子空间语音增强算法。该算法利用带噪的语音信号数据训练一组深度神经网络语音生成型模型(DNN训练模型);在测试增强阶段根据噪声估计和DNN模型去除非平稳噪声;最后,通过信号子空间在抑制噪声和减少信号失真上做出较为折中的选择重构语音信号。实验结果表明,基于深度神经网络的子空间语音增强算法对非平稳噪声有非常强的抑制能力,通过STOI和PESQ值反映了在低信噪比下,该算法可以提高增强语音的可懂度。
The random and abrupt nature of noise makes the traditional algorithm more difficult to suppress non-stationary noise than stationary noise.To solve above problems,a speech enhancement method based on Deep Neural Network and subspace algorithm was proposed.First,the algorithm uses speech signal with noise to train a set of speech production model by deep neural network（DNN training model）.Second,in the test of the enhancement noise estimation and DNN model are used to remove non-stationary noise.Finally,the signal subspace in suppressing noise and reducing the signal distortion makes a more eclectic choice of speech signal reconstruction.The simulation results show that the improved algorithm based on speech enhancement can remove non-stationary noise strongly and improve the intelligibility of enhanced speech by STOI under low SNR.
出处
《太原理工大学学报》
CAS
北大核心
2016年第5期647-650,679,共5页
Journal of Taiyuan University of Technology
基金
国家自然基金项目资助:基于认知机理的情感语音识别基础研究(61370093)
山西省青年科技研究基金资助项目(2013021016-1)
山西省自然科学基金资助项目(2013011016-1)
校基金团队资助项目(2014TD028
2014TD029)
