Combining acoustic features and medical data in deep learning networks for voice pathology classification

In this paper, we present a study on the efficiency of neural networks for the hard problem of automatically classifying voice disorders. To this end, convolutional architectures combined with feed-forward neural networks are used for the classification of four types of voice disorders. Speech signals and data from medical records, collected by the Far Eastern Memorial Hospital (FEMH), involving four speech pathologies, (functional dysphonia, phonotrauma, laryngeal neoplasm and unilateral vocal paralysis), were analyzed and the proposed method participated at the FEMH Voice Data challenge 2019. The respective classification accuracy at the challenge's testing dataset was 57% and the method ranked fifth with a small performance margin from the leading method.