
162
was obtained. When the subset of features consists of the five features, the Hurst parameter,
FMMI, fuzzy entropy, second-order Rényi entropy and F0, the best average recognition rate
(88.66%) was obtained. From the above results, it is suggested that the nonlinear acoustic
characteristic can more fully characterize the pathological voice.
The saliency measure based on the BP neural network can perfectly reduce the feature
dimension and prove the recognition rate. Also, it effectively reflects the individual character-
istic’s contribution to the voice pathology detection. Although there is not much contribution
made by the traditional features ...