半连续HMM码本生成算法的研究

The Research of Codebook Generation Algorithm in SCHMM

对基于半连续隐马尔科夫模型(SCHMM)语音识别系统的码本生成算法及其原理进行了探讨。阐述了译码器扰动简化随机松弛聚类算法(SR-D)，并将其应用到初始码本生成中。实验结果表明这种方法能显著地提高系统性能。初始码本生成后，采用最大似然准则对生成的码本进行了训练，使得码本和SCHMM其它参数达到较好的一致。也探讨了码本大小及其对最终性能的影响并给出了相关实验结果。

This paper discusses the algorithms of codebook generation in SCHMM based speech recognition system and their fundamental principle. Introducing the basic discipline of Stochastic Relaxation- Division (SR-D) algorithm, it introduces the general idea into the authors' initial codebook generation algorithm. The experiment results show that the new method significantly improves the system performance. With the modified initial codebook, the model is further trained with the criterion of Maximum Likelihood(ML),which leads the codebook and other parameters more consistent and compatible.The effect of codebook size to final performance is also discussed and relative experiment results are presented.