Segmentation and accuracy-based scores for the automatic assessment of oral proficiency for proficient L2 speakers

Abstract

This paper reports on the automatic assessment of oral proficiency for advanced second language speakers. A spoken dialogue system is used to guide students through an oral test and to record their answers. Indicators of oral proficiency are automatically derived from the recordings and compared with human ratings of the same data. The proficiency indicators investigated here are based on the temporal properties of the students’ speech as well as the their ability to repeat test prompts accurately. Results indicate that, both for segmentation as well as accuracy-based scores, the most simple scores correlate best with the humans’ opinion on the students’ proficiency. Combining different scores using multiple linear regression leads to marginally higher correlations. However, these improvements are too small to justify the associated increase in the computational complexity of the system.