论文部分内容阅读
为了实现语音识别中基于隐Markov模型(hidden Markov model,HMM)的满方差建模,该文提出了基于树的相关系数的补偿方法。首先自顶向下构建状态的回归树,用简化的仅考虑协方差的对称Kullback-Leibler散度来度量Gauss之间的差异。每个Gauss核接到相应状态下作为叶子节点。叶子节点的相关系数矩阵用其父节点及祖先节点的相关系数矩阵的线性插值得到。线性插值权在最大似然意义下进行优化。实验结果显示取得的识别性能相对异方差线性判别分析、半绑定协方差、基于树的协方差非对角补偿方法的字误识率分别相对下降9.71%、9.17%和4.12%。
In order to realize full-variance modeling based on hidden Markov model (HMM) in speech recognition, this paper proposes a tree-based compensation method of correlation coefficient. First, build the state regression tree from top to bottom, and measure the difference between Gauss with a simplified symmetric Kullback-Leibler divergence that considers only covariance. Each Gauss kernel is connected to the corresponding state as a leaf node. The correlation coefficient matrix of leaf nodes is obtained by linear interpolation of the correlation coefficient matrix of its parent node and its ancestor node. Linear interpolation weights are optimized in the maximum likelihood. Experimental results show that the recognition rate of relative heteroscedasticity linear discriminant analysis, semi-bound covariance, tree-based covariance non-diagonal compensation method of word recognition rate decreased by 9.71%, 9.17% and 4.12% relative decline.