Auditory Feature Extraction Approach for Robust Pathological Voice Recognition

While binary discrimination between healthy and pathological voices provides initial screening capability, clinical utility requires accurate multi-class classification to distinguish between different pathology types. This paper presents a novel Auditory Feature Extraction (AFE) approach, inspired by auditory perception mechanisms, for a robust multi-class pathological voice recognition (PVR) system. The proposed approach uses a Gammachirp FilterBank (GCFB) to accurately simulate the spectral behavior of the cochlea. This FilterBank is considered an efficient auditory filter model and contains 128 Gammachirp auditory filters with center frequencies equally distributed according to the Equivalent Rectangular Bandwidth-rate scale. The GCFB output undergoes decimation, cubic-root amplitude compression, and then Discrete Cosine Transform to finally generate the AFE coefficients. We evaluated the performance of the proposed AFE approach for PVR using Hidden Markov Model Toolkit on two widely recognized datasets: the Saarbruecken Voice Database (SVD) and the Massachusetts Eye and Ear Infirmary (MEEI) dataset. Our AFE outperforms state-of-the-art approaches as revealed by assessment results compared to Human Factor Cepstral Coefficients (HFCC), Frequency Domain Linear Prediction (FDLP), and Mel-Frequency Cepstral Coefficients (MFCC). AFE approach achieves 99.75% balanced accuracy in binary voice classification (95.6%, 94.8%, and 93.85% for HFCC, FDLP, and MFCC, respectively) and 94.38% balanced accuracy in multi-class pathologic voice classification (72.93%, 69.66%, and 60.03% for HFCC, FDLP, and MFCC, respectively) on the SVD dataset, while accomplishing 100% balanced accuracy on the MEEI database. These findings suggest the AFE approach provides a robust and highly discriminative feature set that could lead to an improved voice pathology classification, then a better clinical screening of voice disorders.