Implementation of intelligent signal processing method for vocal fold pathology detection through speech signals

Humans produce voice through lungs and vocal folds or vocal cords in the larynx. The presence of pathologies or diseases in vocal folds could have an impact on the normal vibratory patterns and sound quality. The vocal fold pathology detection was a difficult task because some current diagnosis such as laryngoscopy and video endoscopy are invasive and it needs expert and exact knowledge to analyze the human voice signal parameters. A significant non-invasive diagnostic technique based on voice signal processing i.e. an automated vocal fold pathology detection system to find various parameters or features obtained from the voice signal is proposed. This method is non-invasive and can be useful as detection tool to assist doctors to make better decisions on patients' conditions for further clinical examinations. Most researchers used features in frequency-domain to find the voice parameters. A few limitations were discovered such as the loss of time-domain information while performing the frequency transformation making it difficult to define and diagnose specific voice disorders clinically. Also, majority of works are limited to pairwise classification problems and less focus on the classification of the exact pathology. This study investigates the performance of features derived from the Dual-Tree Complex Wavelet Packet Transform (DT-CWPT) with energy and entropies measures quantified with two classifiers, k-Nearest Neighbours (k-NN) and Support Vector Machine (SVM). The DT-CWPT introduces the complex coefficients in time and frequency scale which delivers a better time-frequency resolution, provides approximately shift-invariance and good directional selectivity where these properties are lacking in traditional wavelet transform. Decomposition is done on the voice signals using energy and entropies to signify the complexity of voice signals in the time and frequency domains. Five sets of datasets obtained from the Massachusetts Eye and Ear Infirmary (MEEI) voice disorders database and Saarbruecken Voice Database (SVD) are used. In order to reduce the high dimensionality feature of multiclass datasets, feature selections using ReliefF algorithm and Genetic algorithm (GA) are proposed to reduce redundancy features and obtain the optimum features for classification. This automatic voice pathologies detection experimented with file-based and frame-based analyses for pairwise (normal and abnormal class) and multiclass classification of the abnormal voice pathologies (Vocal Fold Cysts, Vocal Nodules, Polyp, and Paralysis). It is observed that the best accuracies obtained using a combination of DT-CWPT with SVM classifier for the pairwise analysis achieved 99.73% to 100% accuracy meanwhile, 93.90% to 99.64% of average accuracy achieved in multiclass. The presented study provides a promising detection rate for multiple voice disorders and is useful for the development of computer-based diagnostic tools for voice pathology screening in health care facilities.