基于感知模型的美尔谱失真测度

Mel-Spectral Distortion Measure Based on Perception Model for Objective Speech Quality Assessment

为了有效评价通信系统中的语音质量,基于语音感知分析,提出了M el域上一种新的语音信号特征表示方法———MFSC(美尔谱系数).MFSC既考虑人耳对频率的非线性感知特性,又结合了声音强度-响度非线性变换特性,符合语音感知分析.基于MFSC特征参数的提取,提出了用于语音质量客观评价的美尔谱失真测度(M el-SD),并将其应用于干扰条件下的无线通信系统语音质量评价.实验结果表明,M el-SD的平均相关值为0.942,分别比M el-CD和PESQ(语音质量感知评价)提高了0.089和0.031.

In order to assess speech quality effectively, a new approach of feature extraction of speech signals, MFSC （Mel-frequency spectral coefficient ）, was proposed on the basis of the speech perception model. This approach considers both the perceptual nonlinear relation between sound intensity and loudness and the nonlinear perception for frequency in Mel domain. Furthermore, an objective speech quality measure with MFSC feature parameters, referred to as Mel-SD （Mel-spectral distortion measure）, was put forward, and speech quality assessment based on Mel-SD for jammed wireless communication systems was given. The experimental results show that the average correlation using Mel-SD is 0. 942, increasing by 0. 089 and 0. 031 respectively compared with Mel-CD （ Mel- cepstral distance measure） and PESQ （perceptual evaluation of speech quality）.