We discuss the possibilities of automatic speech recognition (ASR) based software for diagnostic feedback. The aim of our research is to develop personalized ASR so that it can be utilized as both self-learning and assessment tool. This will raise learners´ linguistic awareness and enrich the construct of learner autonomy.
Despite the widely recognized gap between communicative teaching and slightly outdated testing practices, the majority of scholarly publications focus around face-to-face training and assessment of oral proficiency, while research on digitally delivered and automatically scored speaking tasks is scarce. There are, though, promising recent examples of large-scale computerbased speaking tests based on speech recognition technology for several different languages and age-groups (Evanini, Hauck & Hakuta 2017). In this focused multimodal presentation, we introduce a new research project, which builds on these examples. The aim of our project is to develop personalized automatic speech recognition (ASR) so that it can be utilized as both selflearning tool and for formative assessment purposes. The approach is to first train native ASR models for L1 and L2 and then personalize the models to each individual L2 learner. When the system is aware of the student’s native speech phonology, it can follow how the pronunciation gradually adapts from L1 towards L2. By computationally modeling the learning process while the learner speaks, the system is not only able to accurately recognize what the learner says and what mistakes (s)he makes, but also to follow his/her pronunciation learning, continuously assess his/her skills and give accurate, advising feedback. This will raise learners´ linguistic awareness and enrich the construct of learner autonomy, since learners will be able to track the development of their language proficiency and detect their strengths and challenges. The feedback to the L2 learner will be given on phonemes, prosody, fluency, and vocabulary. The focus will lie on features that human raters assess as important for comprehensible speech (Saito et al. 2017). In our presentation, we discuss the possibilities of ASR based software for this kind of diagnostic feedback. Also, we will bring up the challenges we have encountered or presumably will encounter during the project.
