基于BP神经网络的管制语音质量评价方法  

Control Speech Quality Assessment Method Based on BP Neural Network

作　　者：傅强[1] 李贵民 吴岳洲 FU Qiang;LI Gui-min;WU Yue-zhou(Civil Aviation Flight University of China,Guanghan 618000,China)

机构地区：[1]中国民用航空飞行学院,四川广汉618000

出　　处：《航空计算技术》2023年第3期1-5,共5页Aeronautical Computing Technique

基　　金：国家重点研发计划项目资助(2021YFF0603904);中央高校基本科研业务费项目资助(ZJ2022-004);中国民用航空飞行学院面上项目资助(JG2022-06)。

摘　　要：针对当前管制语音缺乏高效、客观、接近人耳感知的质量评价方法,提出了一种基于改进MFCC特征和BP神经网络的管制语音质量评价方法。用Gammatone滤波器代替MFCC算法中的三角形滤波器处理管制语音,该滤波器能体现基底膜尖锐的滤波特性。引入一阶差分和二阶差分提取语音的动态特性,将新的MFCC特征与其一阶和二阶差分融合,形成一种更具代表性的高维度特征。将改进的MFCC特征通过BP神经网络映射到MOS值,从而实现管制语音质量评价。根据1740条真实管制语音数据集的实验结果表明,相较于ITU提出的无参考客观评价方法P.563,文中的Karl Pearson相关系数提高了18.46%,均方误差下降了12.81%。In view of the lack of effective,objective and close to human ear perception′s quality evaluation methods of control speech,a new method based on improved MFCC features and BP neural network was proposed.Firstly,Gammatone filter is used instead of the triangle filter in MFCC algorithm to process the control speech.The filter can reflect the sharp filtering characteristics of the substrate film.Secondly,first-order difference and second-order difference are introduced to extract the dynamic characteristics of speech,and the new MFCC features are fused with first-order and second-order difference to form a more representative high-dimensional feature.Finally,the improved MFCC features are mapped to MOS values by BP neural network,so as to achieve control speech quality evaluation.According to the experimental results of 1740 real control speech data sets,compared with the non-reference objective evaluation method P.563 proposed by ITU,the Karl-Pearson correlation coefficient in this paper is increased by 18.46%and the mean square error is decreased by 12.81%.

关 键 词：MFCC 语音质量 Gammatone BP神经网络 

分 类 号：V355[航空宇航科学与技术—人机与环境工程]