Machine Learning to Detect Stuttering : Detecting Different Types of Stuttering Using Multiple Machine Learning Algorithms

Abstract

Stuttering is an important problem largely addressed with speech therapy. It can adversely affect the quality of the affected person’s daily life. If machine learners can be designed to detect stuttering types in speech, they have the potential for helping affected persons speak more fluently. Recently, there have been some labeled speech datasets made publicly available, such as the Sep-28k and FluencyBank datasets. These datasets collectively account for 32,000 three-second audio clips containing labels on whether stuttering is present (and its types) or not. This work examines five machine learning models, including Decision Trees, Random Forest, Gradient Boosting, XGBoost, and CatBoost as models for stuttering prediction. In addition, different context-relevant empirical case studies are conducted, including predicting an individual stuttering type, predicting different types of stuttering with each machine learner, and predicting the presence or absence of any stuttering speech in an audio clip. The features in the model are a combination of audio features extracted using the Librosa package in Python, and a novel audio feature as proposed in this work. The prediction accuracy of the different models is in the 80% range, which is competitive, or better than, some of the more advanced models in related literature.