结合模型混淆度和BIC准则的语种识别精细建模方法

摘要

提出了一种基于语种模型混淆度的模型参数估计方法,并结合贝叶斯信息准则(Bayesian information criterion,BIC)来进行模型的选取,避免了大量标注信息的需求.在NIST-07语种识别30,10和3s的测试任务中,分别给出了在最大似然(Maximum likelihood,ML)准则和最大互信息(Maximum mutual information,MMI)准则下性能比较,所提出的方法相对于基线系统,性能都有明显的提升,而且达到了利用标注信息进行细化建模相同的水平.%In language identification, the detailed annotations can make the model parameters estimate more accurately, but these annotations are difficult to be acquired. This paper proposes a kind of model parameters estimation method based on language model confusion, combining Bayesian information criterion (BIC) for model selection, avoid acquiring large annotations. In NIST07 language recognition for 30, 10 and 3 s test tasks, we presented performance comparison in the maximum likelihood (ML) criterion and the maximum mutual information (MMI) criterion, our proposed method have significantly improvement, and reach the same level as using annotations.