英语口语机考系统中语种识别方法的研究

A Study on Language Identification in Call English Oral Test

在基于GMM的语种识别系统中,实际环境和个人因素一直是影响识别率提高的因素。从各模型帧似然概率的统计特性出发,提出了一种新的帧似然概率变换方法,它把帧概率按照一定规则变换成权值。识别时不依靠概率的绝对值,而是依靠帧概率在所有其它语种模型中的相对位置来决定模型总得分。理论分析和实验结果表明,同传统方法相比,本文提出的方法能较好地提高语种识别率。

In GMM-based the factors that influence language identification system, the environment and individual characteristics are always the identification accuracy. A new conversion method based on likelihood probability of frames is presented, which analyzes the statistical specialities of likelihood probabilities of frames, and the weights are got from the probabilities of frames according some rules, the recognition result doesn＇t depend on the absolute value of the probability, but the model score is decided by relative positions of the frame probability in all other language models. The theoretical analysis and experimental results show, compared to traditional method, the method presented in the paper can better improve the identification accuracy.