An approach on a combination of higher-order statistics and higher-order differential energy operator for detecting pathological voice with machine learning

Voice signal is an indicator finding a progression of diseases such as nerve disorder and muscle dysfunction. To improve the performance of medical diagnosis system using the voice signal, this paper suggests a new feature extraction method which combines higher-order statistics (HOS) and higher-order differential energy operator (DEO). For the experiment, Saarbruecken Voice Database (SVD) was used, and 687 healthy voice samples and 263 pathological voice samples which consist of Cysts, Paralysis, and Polyp were selected. In addition, the OpenSmile script which provides 6,373 features was used for comparison with our new features. To decide the most effective features, Gradient Boosting was conducted as a feature selector. Finally, 20 features including 15 combinations of HOS and DEO were chosen, and deep neural network(DNN) was trained using the new features. The best accuracy of 87.4% was obtained, which exceeds the best accuracy of 84.5% with the existing features. The finding suggests a possibility that the pathological voice can be efficiently detected with only statistical information without heavy computations such as convolutional neural networks. Due to the simple structure, we expect this approach will be easily applied to a variety of mobile systems.