Can Acoustic Features of Students' Oral Reading Predict Their Oral Reading Fluency? A Machine-Learning Approach

• In Event: Embracing New Assessment Methods to Enhance Measure and Predict Learning Outcomes

Abstract

This study aims to explore the potential of using acoustic features to predict students’ oral reading fluency using machine learning techniques; and compare error patterns of predictions between students with North American accent and those with other regional accents. For 158 grades 4–6 students’ oral reading recordings, 1,581 acoustic features were extracted and used to predict words correct per minute. Machine-predicted score was strongly correlated with human-rated score (r=.91). Although the size of prediction errors was not significantly different between the North-American- and other-regional-accent groups, machine overestimated low-performers while underestimating high-performers, regardless of accent. The findings have promising implications for further use of natural language processing features that can significantly reduce the time spent on administering ORF assessment in classrooms.

Authors

• Hyunah Kim, Education Quality and Accountability Office

• Liam Hannah, University of Toronto

• Eunice Eunhee Jang, University of Toronto