基于神经网络的语音谱失真测度研究  

Research on speech distortion measure based on neural network

作　　者：付强[1] 田斌[1] 张知易[2] 易克初[1] 

机构地区：[1]西安电子科技大学ISN国家重点实验室,西安710071 [2]西南通信研究所,成都610041

出　　处：《声学学报》2001年第2期180-184,共5页Acta Acustica

摘　　要：提出了基于神经网络的语音谱失真测度概念。利用前向神经网络，包括多层感知器和径向基函数网络，对多维非线性函数的逼近原理，使得谱失真测度函数具备了表现人耳听觉系统的主观感知行为的能力。结合语音质量客观评价应用，我们以在大量的失真条件下得到的主观评价结果作为期望值对该网络进行训练。统计相关分析表明，基于神经网络谱失真测度的客观评价方法的主客观评价的相关性，较之传统欧氏距离以及加权欧氏距离都有了显著的提高，并具有更高的鲁棒性．该方法还具有技术独立性．This paper puts forward a novel spectrum distortion measure (MSD) of speech based on Artificial Neural Network (ANN). It provide the MSD with the capability of employing subjective behavior of perception of human's auditory system, by using the theory of approaching to nonlinear function of multilayer perceptrons (MLP) and radial basis function networks (RBFN). In the application of speech quality objective assessment, the result of subjective evaluation under plenty of distortion conditions is directly used as the desired value of training. Statistic analysis shows that, comparing with traditional Euclidian distance and weight Euclidian distance, the correlation of subjective to objective of MSD based on ANN has great improvement, also are the reliability and'robustness. Moreover, it has the property of technology independent.

关 键 词：语音质量客观评价 谱失真测度 多层感知器 径向基函数网络 前向神经网络 欧氏距离 神经网络 语音谱 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]