Sound feature space effects on the performance of annoyance evaluation model based on support vector machine

Abstract

Definition of sound feature space affects the performance of a sound quality evaluation (SQE) model. In this paper, two types of sound feature space, which consider psychoacoustic metrics (the first type) and auditory band sound features (the second type), were discussed in SQE modeling. Three sound feature spaces based on the loudness and sharpness, critical band loudness and auditory feature band energy were presented, respectively. By using the measured interior noises at different vehicle speeds, a jury test was implemented for subjective annoyance evaluations. Based on the extracted feature vectors, the noise sample points were projected in three sound feature spaces, respectively. The mapping relationships between the subjective evaluation result space and the different sound feature spaces were established by using the support vector machine (SVM) algorithm. Three annoyance evaluation models (AEMs), i.e., psychoacoustic metrics based AEM (PM-AEM), critical band loudness based AEM (CBL-AEM) and band energy based AEM (BE-AEM), were developed. The cross-validation results indicated that the prediction accuracies and stabilities of the CBL- and BE-AEMs are improved comparing with that of the PM-AEM. The proposed sound feature space based on critical band loudness and the space in which auditory band energies are base vectors can better represent the annoyance evaluation than the space based on loudness and sharpness. Considering the multi-band filtering property of cochlear and the performances of the PM-, CBL- and BE-AEMs, a conclusion can be drawn that, the second type of sound feature space describing the detailed distribution of the auditory band information is a better selection for annoyance modeling of vehicle noise than the first type of feature space.