A phone verification approach to pronunciation quality assessment for spoken language learning

Abstract

Computer-assisted language learning (CALL) is a form of computer-based assisted learning used in teaching to facilitate the language learning process. One major aspect of CALL for spoken language learning is the automatic assessment of pronunciation quality. It greatly relies on speech recognition technology to provide gradings for the pronunciation quality of the given input speech. This paper introduces a phone verification approach which allows the detection of mispronunciations at phone level. The detection thresholds can be determined based on the equal error rate (EER) metric using a database containing only native speech. In addition this approach also allows aggregation of assessment scores at sentence and speaker levels by computing the average phone rejection rates. This paper compares three different methods of generating goodness of pronunciation confidence scores. In addition, this paper also examines both unsupervised versus supervised adaptation techniques to improve the verification performance. Experimental results are reported based on the EER metric.