Deep Neural Networks for Voice Quality Assessment Based on the GRBAS Scale

Abstract

In the field of voice therapy, perceptual evaluation is widely used by expert listeners as a way to evaluate pathological and normal voice quality. This approach is understandably subjective as it is subject to listeners’ bias which high inter- and intra-listeners variability can be found. As such, research on automatic assessment of pathological voices using a combination of subjective and objective analyses emerged. The present study aimed to develop a complementary automatic assessment system for voice quality based on the well-known GRBAS scale by using a battery of multidimensional acoustical measures through Deep Neural Networks. A total of 44 dimensionality parameters including Mel-frequency Cepstral Coefficients, Smoothed Cepstral Peak Prominence and Long-Term Average Spectrum was adopted. In addition, the state-of-the-art automatic assessment system based on Modulation Spectrum (MS) features and GMM classifiers was used as comparison system. The classification results using the proposed method revealed a moderate correlation with subjective GRBAS scores of dysphonic severity, and yielded a better performance than MS-GMM system, with the best accuracy around 81.53%. The findings indicate that such assessment system can be used as an appropriate evaluation tool in determining the presence and severity of voice disorders.