Predicting judged fluency of consecutive interpreting from acoustic measures: Potential for automatic assessment and pedagogic implications

Abstract

This experimental study of consecutive interpreting investigates whether: (1) there is any correlation between assessments of its fluency and accuracy; (2) judged fluency can be predicted from computer-based measurements like articulation rate. Ten raters judged six criteria of accuracy and fluency in two consecutive interpretations of the same recorded source speech, from Chinese A into English B, by 12 trainee interpreters (seven undergraduates, five MA students). The recorded interpretations were examined with the speech analysis tool PRAAT. From a computerized count of the pauses thus detected, together with disfluencies identified by raters, 12 acoustic measures of fluency were calculated. The advanced students were more fluent than the beginners; both groups were less fluent in the initial interpretation. Statistical analysis shows: (1) a strong positive correlation between judged accuracy and judged fluency; (2) strong correlations between judged fluency and objective fluency variables; (3) the usefulness of effective speech rate (number of syllables, excluding disfluencies, divided by total duration of speech production and pauses) as a predictor of judged fluency. Other important determinants of judged fluency were the number of filled pauses, articulation rate, and mean length of pause. Potential for developing automatic fluency assessment in consecutive interpreting is discussed, as are possible training implications.