The accurate and interpretable detection of pathological voice conditions remains a critical challenge in clinical speech analysis. Existing acoustic and deep learning–based approaches are majorly based on spectral or temporal representations that provide limited insight into the underlying physiological behavior of the vocal folds. This work introduces a novel and accurate Recurrence Plot (RP) –based framework that models sustained vowel phonations as nonlinear dynamical systems. This enables direct characterization of vocal fold oscillatory patterns through Recurrence Plots and Recurrence Quantification Analysis (RQA). The proposed method facilitates robust multiclass classification of three common voice disorders such as hyperfunctional dysphonia, laryngitis, and vocal fold polyp. Using the publicly available Saarbrücken Voice Database (SVD), the proposed method achieves a mean classification accuracy of 90.2% and shows superior disease-specific performance. Compared to a state-of-the-art CNN-based temporal feature method, the proposed work showed a relative improvement of 32% for laryngitis and a 32.4% improvement for Vocal Fold polyp, while the performance for hyperfunctional Dysphonia remains comparable. These improvements indicate that modeling speech as a nonlinear dynamical process provides richer and more discriminative representations than traditional linear methods. Thus, by integrating physiological interpretability and strong diagnostic performance, this study presents a clinically meaningful and explainable solution for pathological voice assessment
Krishnan et al. (Thu,) studied this question.