Automatic assessment of children speech to support language learning

Viac o knihe

This work focuses on the pattern recognition aspects of computer-assisted pronunciation training (CAPT) for second language learning. It highlights that while pronunciation training is an emerging area within computer-assisted language learning, it has been addressed only minimally by commercial systems. The thesis extends and combines various approaches, developing a comprehensive set of nearly 200 pronunciation and prosodic features. Pronunciation scoring is treated as a classification task within a high-dimensional feature space, relying on automatic speech recognition as the foundation for most algorithms. A system tailored for school-aged children is introduced, addressing the challenges of recognizing young speakers, who are often poorly understood by existing automatic systems. The system incorporates phonetically motivated rules to handle common mispronunciation errors, enhancing its effectiveness for pronunciation scoring. Evaluating pronunciation assessment algorithms is complex, as there is no objective ground truth; this is illustrated through an analysis of annotations from 14 teachers. However, various measures indicate that the system's accuracy aligns closely with teacher assessments, demonstrating a level of agreement comparable to that among the teachers themselves. The evaluation focuses on native German speakers learning English.