关键词: 高斯混合模型, 病态嗓音, 计盒维数, 截距

Abstract: By analyzing the mechanism of pronunciation, normal and pathological voice of traditional acoustic parameters：fundamental frequency, formant, Mel Frequency Cepstrum Coefficient（MFCC）, and non-linear feature parameters：box-counting dimension and intercept, are extracted as feature vectors of recognition of pathological voice. 156 normal voice samples and 146 pathological voice samples are recognized based on Gaussian Mixture Model（GMM）. The results show that the nonlinear feature parameters of box-counting dimension and intercept can well distinguish between normal and pathological voice. The combination of box-counting dimension, intercept and the traditional acoustic parameters-fundamental frequency and formant can achieve a better recognition rate of 92.60%.

Key words: Gaussian Mixture Model（GMM）, pathological voice, box-counting dimension, intercept