MACHINE LEARNING ANALYSIS OF SPEECH AND EGG FOR THE DIAGNOSIS OF VOICE PATHOLOGY

Abstract

Current approaches to voice diagnosis involve a clinician examining the patient, listening to their voice and in some cases, using additional measurements of the larynx such as EGG. Here we train a feedforward convolutional neural network on a database of normal healthy drama students recorded speaking passages in English, to reconstruct the associated EGG (Lx) waveform. We then use the network to predict the EGG from the acoustic speech signal on a different set of speakers, including ones that exhibit laryngeal pathologies. We show the predicted EGG is very similar to the actual recorded EGG and, as such, can provide a useful indication of voice pathology. Importantly, the network is able predict the pathological EGG waveforms even though it was never trained on pathological speech.