论文部分内容阅读
提出一种称为“受限线性搜索”的优化方法,并用于语音识别中混合高斯的连续密度隐马尔科夫(CDHMM)模型的区分性训练.该方法可用于优化基于最大互信息(MMI)准则的区分性训练目标函数.在该方法中,首先把隐马尔科夫模型(HMM)的区分性训练问题看成一个受限的优化问题,并利用模型间的KL度量作为优化过程中的一个限制.再基于线性搜索的思想,指出通过限制更新前后模型间的KL度量,可将HMM的参数表示成一种简单的二次形式.该方法可用于优化混合高斯CDHMM模型中的任何参数,包括均值、协方差矩阵、高斯权重等.将该方法分别用于中英文两个标准语音识别任务上,包括英文TIDIGITS数据库和中文863数据库.实验结果表明,该方法相对传统的扩展Baum-Welch方法在识别性能和收敛特性上都取得一致提升.
An optimization method called “limited linear search ” is proposed and used to discriminate the hybrid Gaussian Continuous Density Hidden Markov Model (CDHMM) in speech recognition.This method can be used to optimize the algorithm based on maximum mutual information (MMI) .In this method, we firstly consider the Hidden Markov Model (HMM) discriminative training problem as a constrained optimization problem and use the KL metric between models as the optimization process Based on the idea of linear search, it is pointed out that HMM parameters can be represented as a simple quadratic form by limiting the KL metric between the pre-and post-update models.This method can be used to optimize any parameter in the mixed Gaussian CDHMM model, Including mean, covariance matrix, Gaussian weight, etc. The proposed method is applied to both Chinese and English standard speech recognition tasks respectively, including English TIDIGITS database and Chinese 863 database.Experimental results show that the proposed method is superior to the traditional Baum-Welch method Uniform improvement in recognition performance and convergence characteristics.