Automatic Voice Disorder Detection Using Self-Supervised Representations

Many speech features and models, including Deep Neural Networks (DNN), are used for classification tasks between healthy pathological with the Saarbruecken Voice Database (SVD). However, accuracy values of 80.71% phrases or 82.8% vowels /aiu/ highest reported audio samples in SVD when evaluation includes wide amount pathologies database, instead a selection some pathologies. This paper targets this top performance state-of-the-art Automatic Disorder Detection (AVDD) systems. In framework DNN-based AVDD system we study capability Self-Supervised (SS) representation learning describing discriminative cues speech. The processes SS temporal sequence single feed-forward layer Class-Token (CT) Transformer obtaining Furthermore, there is evaluated suitable data extension training set out-of-domain also to deal low availability using models voice pathology detection. Experimental results corresponding dataset, all available, show until 93.36%. means that proposed achieved improvements 4.1% without extension, 15.62% after compared baseline system. Beyond novelty representations AVDD, fact accuracies over 90% these conditions whole milestone disorder-related research. on in-domain related guidance preparation stage. Lessons learned work suggest guidelines taking advantage DNN, boost developing automatic systems diagnosis, treatment, monitoring