Clinically Interpretable MDVP Feature Learning for Robust Pathological Voice Classification Using Stacked and Kernel Extreme Learning Machines

Abstract

Pathological voice disorders are associated with abnormal oscillation of the vocal folds and changes in phonatory traits that can be quantified through acoustic metrics. Although machine learning and deep learning methods have demonstrated promising results in automated voice pathology identification, the application of clinically interpretable acoustic features along with computation-efficient nonlinear learning remains a significant research area. The present study proposes a clinically interpretable feature learning framework for Multi-Dimensional Voice Program (MDVP) via Stacked Extreme Learning Machine (SELM) and Kernel Extreme Learning Machine (KELM) to classify pathological voices. From speech recordings, an extensive depiction of vocal performance is derived to extract parameters of the Multi-Dimensional Voice Program (MDVP) including fundamental frequency, frequency perturbation, amplitude perturbation and nonlinear traits. The derived features are standardized and selected to reduce redundancy and improve discriminative representation. Two complementary learning frameworks are then applied: SELM and KELM are used to model the hierarchical and nonlinear associations in the MDVP feature domain respectively. The results of the complementary classification are fused using a decision-level fusion approach for the final pathological voice classification. The proposed framework is evaluated independently on the Saarbrücken Voice Database (SVD) and the VOICED database, so that no single database is relied upon and evaluations can be performed under different recording and speaker conditions. In addition, ablation experiments are conducted to evaluate the separate contributions of traditional ELM, KELM, SELM and the proposed fused architecture.