Detecting Structural Events for Assessing Non-Native Speech

Structural events, (i.e., the structure of clauses and disfluencies) in spontaneous speech, are important components of human speaking and have been used to measure language development. However, they have not been actively used in automated speech assessment research. Given the recent substantial progress on automated structural event detection on spontaneous speech, we investigated the detection of clause boundaries and interruption points of edit disfluencies on transcriptions of non-native speech data and extracted features from the detected events for speech assessment. Compared to features computed on human-annotated events, the features computed on machine-generated events show promising correlations to holistic scores that reflect speaking proficiency levels.