Pronunciation education system based on the systematization of non-mothor tongue speech prosody using generation process model and speech synthesis

点击翻译按钮获取中文摘要

Fundamental frequency (F0) contours of speech by natives and learners are analyzed using the generation process model. Several findings, such as phrase components being less affected by language differences, are shown. As for utterances by learners, influence of their mother tongue is observed. Since learners utterances involve F0 movements not observable in natives utterances, accent type identifier trained using native s utterances does not work well. To solve this problem, a series of perceptual experiments is conducted using synthetic speech with systematic control on F0 (points of F0 movements, slope coefficients). Based on the result, a threshold method of high-low decision of F0 is developed. Also, generation process model constraints are applied to HMM-based speech synthesis resulting in speech quality improvement. A pronunciation training system on Japanese accent type is developed and evaluated.