Automatic pronunciation assessment and diagnosis is an important part of many computer-aided language learning systems and computer-assisted language testing systems. This paper carries out researches on pronunciation content verification, automatic pronunciation error diagnosis and automatic assessment on global pronunciation quality. The work of this paper mainly includes the following contributions: 1.Proposed a pronunciation content verification method based on phoneme confusion, which was applied in the pronunciation error diagnosis for isolated words and phrases. The new method used pronunciation variation force alignment to carry out phoneme recognition and computed confidence mesure by using phoneme confusion matrix. The experimental results showed that this method not only can effectively play the system “firewall” role, but also is superior to the mainstream method based on posterior probability in terms of speed and flexibility. 2.Against the disadvavantages of the mainstream method based on posterior probability, methods based on multi-knowledge sources were proposed. Firstly, the method based on pronunciation space was introduced, which used multi-dimensional posterior probability features to describe pronunciation space of users, taking full use of the confusion knowledge of phoneme models. Secondly, introduced prior knowledge of pronunciation errors to get modified multi-dimensional posterior probability features and proposed a new method based on restricted pronunciation space for error diagnosis. Lastly, CMLLR speaker adaptation was introduced, which took full use of prior speaker knowledge of users. The experimental results showed that the above approaches can effectively improve the performance of pronunciation error diagnosis. 3.Built a system to carry out automatic pronunciation quality assessment. The system was for paragraph reading, including the speech-recognition-based pre-processing module, the feature extraction module, and the assessment module. In the feature extraction module, the word match score (WMS) and the segment match score (SMS) are used for assessing the integrity of the content. The experimental results showed that the two kinds of features played very important roles in the assessment for paragraph reading. The system was comparable with human raters in overall performances and had been used in actual spoken evaluation projects.
修改评论