Deep learning-based automatic pronunciation assessment for second language learners

Abstract

For second language learners, computer-aided language learning (CALL) is of high importance. In recent years, the use of smart phones, tablets, and laptops has become increasingly popular; with this change, more people can use CALL to learn a second language. In CALL, automatic pronunciation assessment can be applied to provide feedback to teachers regarding the efficiency of teaching approaches. Furthermore, with automatic pronunciation assessment, students can monitor their language skills and improvements over time while using the system. In the current study, a text-independent method for pronunciation assessment based on deep neural networks (DNNs) is proposed and evaluated. In the proposed method, only acoustic features are applied, and native acoustic models and teachers’ reference speech are not required. The method was evaluated using speech from a large number of Japanese students who studied English as a second language.