Canary Speech Fatigue Model for Japanese

Abstract: Toward an automatic health monitoring tool, we investigate voice analysis technology to extract related features from speech and to build a machine learning model for identifying fatigue. We collect voice data and their fatigue labels through phone calls and then experiment with diverse machine learning methods using various acoustic and prosodic features. The models are trained on spontaneous Japanese speech from participants who are older than 70 years. Each model and feature shows different performance and the logistic regression model using x-vectors outperforms other models. This study shows that our proposed machine learning model can identify fatigue by analyzing voice and speech with sensitivity at 0.87 and specificity at 0.65.