【摘要】 在基于隐马尔柯夫模型 (HMM)的语音识别系统中 ,模型训练最常用的算法是 Baum-Welch算法 .该算法具有快速收敛及保证每步迭代模型的似然概率单调增的优点 .但它基于最大似然训练准则 ,而该准则不能将各个模型很好地分开 .这直接导致了识别时的错误 .鉴于最大互信息可以克服这一缺点 ,提出了一种基于最大互信息的训练方法 .该方法借助于梯度 ,调整参数以使模型与训练数据的互信息最大 .实验结果表明 ,使用该方法使系统的识别性能得到了一定的改善更多还原

【Abstract】 In HMM based speech recognition systems, the most commonly used training algorithm for models is Baum Welch algorithm. Its strength lies in fast convergence and monotonous improvement of the likelihood probability. However, it is based on maximum likelihood training criterion and this criterion cannot separate models well. This leads to many errors in recognition directly. Since the maximum mutual information is noted for better discrimination, this paper proposed a training algorithm based on it. With gradient ascent method, the parameters are modified to maximize the mutual information. The experimental results indicate that using this method improves the performance of the system.更多还原