病态嗓声识别特征参数的优化选择

Optimization and Selection of Feature Parameters in Dysphonia Recognition

为提高病态嗓声识别效率,本研究首次采用主分量分析方法对目前常用的27个嗓声特征参数进行了优化分析,考察了仅用少数主分量参数的识别效果及其分量数对结果的影响;同时根据参数对病态嗓声信息敏感程度,使用正交试验法优选出9个较优特征参数,其识别率即可达到原27个参数的识别结果。经两种方法对参数进行优选后识别率分别达到97.23%和98.10%,显著高于未经优选、使用全部27个参数的92.10%识别率。研究结果表明:原27个参数中,2/3的参数不能很好地反映嗓声的病态变化特征,使用优选的含有大量病态嗓声信息的少量特征参数即可大幅提高病态嗓声识别效率。

In order to improve the performance of dysphonia recognition, 27 common used original acoustic feature parameters were optimized and analyzed by using the method of primary components analysis （PCA）. The performances of PCA with only a few primary components and influenced results with changing component numbers were investigated. In this essay 9 optimum feature parameters were selected from 27 original parameters by using orthogonal layout based on their sensitivities in reflecting the voice disorder information, and then nearly the same corrective classification rate of dysphonla recognition was obtained. The optimized rates given by two methods of primary component analysis and orthogonal layout we are 97.23% and 98.10% respectively, much higher than the rate of 92.10% which was got by using the total 27 original acoustic feature parameters without optimization. The results showed that two thirds of 27 original feature parameters could not reflect the main voice disorder characteristics. The higher performance of dysphonia recognition was achieved by using fewer optimized feature parameters which contains the most information of voice disorders.