An anti-noise performance comparison between acoustic features in detecting voice pathology using machine learning

Abstract

Developing communication robots requires to analyze human voice including various kinds of human biological information because the nonverbal information plays an important role in smooth communications between humans and robots. To analyze numerous voices available via the robot by using machine learning, we should take consideration of the existence of noises added to the voices. However, some acoustic features used for sensing human biological information is not designed for the noises. To validate the variation of the accuracy of classification when the voices includes the noises, we compare the classifications using voice indexes proposed for voice pathology estimation and using Mel-Frequency Cepstrum Coefficients(MFCC) in the classification problem of voice pathology as an early study. Experimental results show that classification using MFCC can detect voice pathology more precisely despite the noises while other voice indexes are adversely affected by the noises.