基于相关维数的病变连续语音检测算法

针对人为设定最优采样延迟不能客观反映信号采样延迟和固定相关维数不易描述病变异常语音复杂性的缺陷,文中提出一种基于相关维数的病变连续语音检测算法.该算法在语音信号合理采样延迟区间内不断调整采样延迟,搜索使正常语音与病变连续语音的区分等错误率达到最小的嵌入相关维数,以避免设定采样延迟的缺陷.同时,通过将相关维数曲线划分成子区间,并判定子区间的稳定性,以达到不固定嵌入相关维数的目的.最后,对每个合理采样延迟时间内获取的训练语音的最优相关维数进行等错误率分析,选用具有最小等错误率的相关维数及对应的采样延迟为文中混沌参数,为测试语音提取混沌指数进行正异常区分.实验结果表明,该算法的区分正确率为75.6%,分别比GMM-SVM、Shimmer、固定相关维和采样延迟法、SHR算法和Jitter算法提高7.8%、9.3%、16.0%、18.0%和20.4%.

Detection Algorithm of Pathological Continuous Speech Based on Correlation Dimension

As the pre-set optimal sampling delay cannot objectively reflect the signal sampling delay and the fixed correlation dimension is inefficient in describing the complexity of pathological abnormal speech,a detection algorithm of pathological continuous speech is proposed based on correlation dimension(CD).In this algorithm,to avoid the defects of pre-set sampling delay,the sampling delay is continuously adjusted within a proper range of sampling delay,and an embedded CD is searched to obtain the minimum equal error rate(EER) of normal and abnormal speech discrimination.At the same time,the correlation dimension curve is divided into several sub-intervals,and the stability of the sub-intervals is determined to overcome the drawbacks of the fixed embedded correlation dimension.After the EER analysis of the optimal CD sets of the training speech data acquired in a reasonable sampling delay range,the CD with the minimum EER and the corresponding sampling delay are chosen as the chaotic parameters of the proposed algorithm to perform a discrimination of normal and abnormal speeches.Experimental results show that the proposed algorithm possesses a correct classification rate of 75.6%,which is respectively 7.8%,9.3%,16.0%,18.0% and 20.4% higher than those of the GMM-SVM algorithm,the Shimmer algorithm,the fixed CD-sampling delay algorithm,the SHR algorithm,and the Jitter algorithm.