A Wearable System for Diagnosing and Monitoring the Intensity of Stuttering

We present a battery-powered wearable system that is able to identify the three basic types of speech disfluencies found in people who stutter: blocks, prolongations, and repetitions. Such a system could be used to aid speech pathology clinicians by performing automated diagnosis of stuttering or monitoring the progress of speech therapy, tasks that are currently time-consuming and produce potentially unreliable results. The system uses a deep learning model trained on the SEP-28k dataset and deployed on a microcontroller. It performs speech audio acquisition and model inference in real time and stores the inference results to non-volatile memory. Once stored, the results can be further analyzed on a PC and presented to the clinician. Our deep learning model achieved a classification accuracy of 65%, 71%, and 64% for blocks, prolongations, and repetitions, respectively. We discuss the possible applications of this system in speech disorder diagnosis and therapy as well as potential improvements.