Improvement of Automatic English Pronunciation Assessment with Small Number of Utterances Using Sentence Speakability

The current Computer-Assisted Pronunciation Training (CAPT) system uses DNN-based speech recognition results to evaluate learner’s pronunciation with high accuracy when using many utterances for the evaluation. However, when we use only a few utterances, the accuracy of the CAPT system deteriorates. One reason for the deterioration is that the score calculated by a CAPT system is biased depending on the pronunciation difficulty of the sentences when using a small number of utterances. In this study, we developed a CAPT system that takes the sentence speakability (pronunciation difficulty of sentences) into account. As a result, the correlation coefficient between the human evaluation and the machine score was 0.46 in the conventional method, while it improved to 0.57 with the proposed method.