The BMBF-funded project Aisla aims to be the first to combine interactivity and adaptivity into a single Dialog-Based Intelligent Computer-Assisted Language Learning (DB-ICALL) application. To achieve this goal, it is crucial that Aisla accurately processes the learner’s spoken utterances. While there have been rapid advancements in automated speech recognition (ASR) in recent years, concerns remain about accuracy and bias, especially for non-standard varieties of English, including non-native language. An earlier debate about the use of ASRs in ICALL is worth revisiting with state-of-the-art commercial ASRs and more open tasks such as those found in Aisla. In this project, we aim to investigate to what extent there are disparities in the performance of commercial ASRs between native and non-native speakers of English and how performance may be impacted by grammatical errors typical of learner language. At the level of the Aisla project, the results would either strengthen our confidence in the use of a commercial ASR or apprise us of limitations and necessary adjustments. On a broader level, the findings would inform researchers about the feasibility of using commercial ASRs for ICALL systems targeting spoken language, especially those seeking to provide corrective feedback on grammar proficiency in a more authentic format.
Project members:
Elizabeth Bear, Xiaobin Chen, Steve Bodnar