Application of Structural Events Detected on ASR Outputs for Automated Speaking Assessment

We investigated features reflecting utterance structure and disfluency profile to improve the automated scoring of spontaneous speech responses by non-native speakers of English. Features derived from structural events (SEs), e.g., clause structure and disfluencies, showed promisingly high correlations to the human proficiency scores. However, previous studies were based on speech transcriptions, and the usefulness of these SEderived features on ASR hypotheses was still unknown. In this paper, we reported our studies related to the detection of SEs from noisy ASR outputs and the application of the detected SEs for automated speech scoring. We found that clause boundary (CB) detection was impacted much less than disfluency interruption point (IP) detection when facing ASR errors. Next, several features derived from the SEs were evaluated based on their correlations to human scores and their relative importance in a linear regression model.