Vocal cords diseases detection by multi-band nonlinear analysis of voice

In order to improve the recognition rate of pathological voices caused by disease of vocal cords, multi-band nonlinear analysis is proposed. Gammatone filter bank is applied to voice signal for front-end time-domain filtering, and then calculate the largest Lyapunov exponent of every band. Data is first mapped into kernel space and use Gaussian maximum likelihood rule to get the best parameter for kernel, which is used for kernel principal component analysis to extract feature. The proposed feature achieves higher recognition rate of 6.25% and 8.45% than MFCC and the largest Lyapunov exponent respectively. When the proposed kernel function is used for kernel principal component analysis, it achieves better performance than traditional function. Ultimately, we get recognition rate of 97.82% by combing them.