A Scale Invariant Technique For Detection Of Voice Disorders Using Modified Mellin Transform

The development of non-invasively derived measures for accurately detecting pathologies that can remove practitioner subjectivity and can save time and cost is an ongoing goal in the field of clinical voice assessment. In this paper the disordered voice samples are detected by deriving Modified Mellin Transform of Log Spectrum (MMTLS) of the voice signals as the feature. Mellin Transform is an integral transform with close relation to the theory of Dirichlet series. The effect of variation in the vocal tract length with gender and age can cause unpredictable results in voice pathology detection. MMTLS is an algorithm which is scale invariant and hence provides a discriminative feature irrespective of gender and age. The work can also be used for detecting disorders from persons with outlier vocal tract lengths. Experiments performed on the Massachusetts Eye and Ear Infirmary (MEEI), Voice and Speech Lab's Disordered Voice Dataset (Dataset 1) and a locally prepared dataset (Dataset 2), results in classification accuracies of 96.48%, 95.92% respectively between normal and pathological voices using an Artificial Neural Network (ANN) with 10 hidden layers.