论文部分内容阅读
研究隐马尔可夫模型 (HMM)的一种有区分力的训练方法 .在多层前向神经网络的框架中实现了 HMM的前向概率计算 .基于这一框架 ,利用偏导数的反向传播计算方法 ,通过梯度上升的优化过程来实现互信息的最大化 ,从而对 HMM进行有区分力的训练 .这一训练方法被称之为 HMM的反向传播训练方法 .此外 ,还设计了一个用以实现这一训练方法的在数值计算上具有强鲁棒性的算法 .语音识别的实验结果证实了这一训练方法的优越性 .
A discriminative training method for Hidden Markov Models (HMM) is studied. The forward probability calculation of HMM is implemented in the framework of multi-layer forward neural network. Based on this framework, the use of partial derivative backpropagation Calculation method, through the process of gradient ascending optimization to maximize the mutual information, and thus differentiate the HMM training.This training method is called HMM back propagation training method.In addition, also designed a In order to realize the computationally robust algorithm of this training method.Experimental results of speech recognition demonstrate the superiority of this training method.