Detection of Vocal Nodules from Phonation Data Using Spectral Features and Recurrence Plots

Vocal nodules are one of the most common forms of voice pathology which, if neglected, can grow and even become cancerous. Diagnosis of vocal nodules involves video stroboscopy in hospital conditions, and is often missed in the early stages which are treatable without surgery. Hence, simple methods based on phonation have gained prominence in diagnosis of voice pathologies. In this work, a novel feature extraction method based on a combination of MFCC and nonlinear recurrence plots is proposed. To compare the performance of this method, comparative results on a standard public dataset, and on the first clinically acquired dataset on an Indian population are presented. The proposed method yields a best accuracy of over 95% on the SVD as well as on the hospital-acquired dataset. The algorithm correctly classifies several early nodule data present in SRIHER data. It is shown that the misclassification occurs due to a few specific samples in the dataset. This work shows that recurrence plots can be a valuable additional feature for population-agnostic accurate detection of vocal nodules.