Training a CNN to Estimate Voice Pathology from Connected Speech Using EGG to Automatically Label the Dataset for Voicing

Abstract

We describe a new system for estimating voice pathology directly from the acoustic speech signal to assist in the diagnosis of pathological voice conditions by voice specialists. Our main novel contributions are the use of Electroglottography (EGG) in neural net training to automatically label speech acoustic signals for voicing and the generation of running estimates of pathology with high temporal resolution from the acoustic signal alone. These estimates can also be linked to the parts of speech signals where voice pathology manifests itself most strongly. By operating directly on the acoustic signal waveform without the use of any pre-processing, we avoid the use of hand-crafted features. We trained and tested a neural network using speech datasets with normal and pathological voicing and found that it can provide effective finegrained indications of pathology. Our quantitative results show that this neural network performs well in distinguishing between speakers with normal and pathological voice conditions, achieving a recognition rate of 91%, which compares favorably with results from other studies.