量化子空间分布隐马尔可夫模型的间接训练

研究了量化子空间分布隐马尔可夫模型(quantized subspace distribution hidden markov model，QSDHMM)间接训练所涉及的三个关键的问题：连续分布隐马尔可夫模型(continuous distribution hidden markov model，CDHMM)的训练、特征子空间的划分和子空间高斯分布的量化方案。提出了相关特征子空间的定义及实现算法。在仿真实验中，采用基于分类学的：Bhattacharyya距离测度的K均值高斯聚类算法对子空间高斯分布进行量化，用最相关子空间的概念划分特征子空间，并将这两者结合使用，提高了系统的识别精度和速度。

Indirect Training of Quantized Subspace Distribution Hidden Markov Model

Recently, we propose a new acoustic model which we call quantized subspace distribution hidden Markov model (QSDHMM), this paper investigate three key problem about indirect training of QSDHMM.We present a definition for subspace in which the features are most correlation and derive a modified K means Gaussian clustering algorithm using the Bhattharyya distance.Combined the new algorithm with the new subspace definition,we obtain significant reduction of the total computation time and memory size in our computer simulation.