Kuo Wang ^1* Xin Qiao ² George Sammit ³ Eric Larson ⁴ Joseph Nese ⁵ Akihito Kamata ¹

• ¹ Simmons School of Education & Human Development, Southern Methodist University, Dallas, United States
• ² University of South Florida, Tampa, Florida, United States
• ³ Southwest Research Institute (SwRI), San Antonio, Texas, United States
• ⁴ Department of Computer Science, Southern Methodist University, Dallas, United States
• ⁵ University of Oregon, Eugene, Oregon, United States

Automated assessing prosody of oral reading fluency presents challenges due to the inherent difficulty of quantifying prosody. This study proposed and evaluated an approach focusing on specific prosodic features using a deep-learning neural network. The current work focuses on crossdomain performance, researching how generalizable the prosody scoring is across students and text passages. The results demonstrated that the model with selected prosodic features had better crossdomain performance with an accuracy of 62.5% compared to 57% from the previous research. Our findings also indicate that students' reading patterns influence cross-domain performance more than specific text passage patterns. In other words, letting the student read at least one passage is more important than having others read all passage texts. The specific prosodic features had a high generalization to capture the typical prosody characteristics for achieving a satisfactorily high accuracy and classification agreement rate. This result provides valuable information for developing future automated scoring algorithms of prosody. This study is an essential demonstration of estimating the prosody score using fewer selected features, which would be more efficient and interpretable.