It is unclear whether corrective feedback (CF) provided by L2 computer-assisted pronunciation training (CAPT) tools must be 100% accurate to promote an acceptable level of improvement in pronunciation. Using a web-based interface, 30 native speakers of Chinese completed a pretest, a computer-based training session to produce nine sound contrasts in English, and a posttest. The study manipulated feedback accuracy using a modified “Wizard of Oz” protocol in which a phonetically-trained human listener in a separate room provided CF on the trainees’ productions, but the trainees thought that the computer-based system provided the CF. The computer system presented a set of three sound contrasts with 100% accuracy, three with 66% accuracy (with one of three human responses changed randomly), and three with 33% accuracy (with two of three human feedback responses being changed). The trainees’ pre- and posttest productions were rated for accuracy by native speakers of English. For trained items, productions were not significantly different when the trainees received CF with 100% or 66% accuracy, but both resulted in greater improvement than feedback with 33% accuracy. An important implication for L2 pronunciation training software is that machine feedback can be beneficial even when it is ‘good enough’ (i.e., not 100% accurate).
Silpachai et al. (Mon,) studied this question.