Abstract: This paper presents a new algorithm which can assess the pronunciation quality of the English spoken by Chinese students. The new algorithm uses Gaussian Mixture Model(GMM) and Universal Background Model(UBM), which is successfully used in speaker verification. It calculates the duration normalized log-likelihood ratio of each phone as phonemic pronunciation scores. It combines each phonemic score to obtain the overall pronunciation quality. The algorithm is evaluated by using a corpus of non-native speech. Experimental results show that the approach outperforms other assessment algorithms on correlations with expert scores at the sentence level. In the test database, this method obtains high correlation(0.700).

Key words: Universal Background Model(UBM), log-likelihood ratio, Gaussian Mixture Model(GMM), pronunciation quality scoring

摘要： 将已经成功应用到说话人识别/确认领域中的高斯混合模型和全局背景模型(UBM)引入语音发音质量评价领域，提出一种新的评价英语发音质量的算法。该算法训练出标准发音的全局背景模型。UBM模型描述与音素无关的特征分布，定义段时长归一化的相似度比例对数为音素的发音质量分数，综合得到整句发音的评分结果。实验证明，在实验室自行采集的非母语语音数据库上，该算法评分与专家评分的相关性达到了0.700，优于其他评分算法。

关键词: 全局背景模型, 对数似然比, 高斯混合模型, 发音质量评价

CLC Number: 

• TP312