摘要
隐Ｍａｒｋｏｖ模型（离散ＨＭＭ）的参数估计问题，是ＨＭＭ在语音处理应用中的关键问题。经典的Ｂａｕｍ＿Ｗｅｌｃｈ算法是基于最陡梯度下降的局部优化算法，ＨＭ Ｍ模型的质量取决于初始模型的设计。解决这一问题的根本方法在于使算法具有随机性。本文结合随机松弛算法（ＳＲ）的全局搜索能力和Ｂａｕｍ＿Ｗｅｌｃｈ算法的局部优化性能，提出了一种离散隐 Ｍａｒｋｏｖ模型参数的全局优化算法。该算法根据 ＨＭＭ的参数对 Ｐ（Ｏ／λ）的不同影响，对观察值概率矩阵Ｂ进行满足一定降温规范的随机扰动，可对离散ＨＭＭ的参数进行全局优化训练。
The training of HMM is the key problem for speech processing. Because the application of conventional Baum_Welch algorithm tends to arrive at a local optimization, a global optimization algorithm based on stochastic relaxation(SR) algorithm is proposed. According to the different influence of each parameter upon P(O/λ), a stochastic perturbation satisfying some temperature specification is added to the observation value probability matrix B to optimize the discrete HMM parameter globally.
出处
《电路与系统学报》
CSCD
2000年第3期78-81,共4页
Journal of Circuits and Systems
基金
国家自然科学基金!(69872036)
关键词
离散HMM
全局优化
MARKOV模型
语音识别
discrete Hidden Markov models
stochastic relaxation
global optimization
Baum_Welch algorithm,
