基于隐马尔可夫模型与并行模型组合的特征补偿算法  

Feature compensation algorithm based on hidden Markov model and parallel model combination

作　　者：吕勇[1] 吴镇扬[1] 

机构地区：[1]东南大学信息科学与工程学院,南京210096

出　　处：《东南大学学报（自然科学版）》2009年第5期889-893,共5页Journal of Southeast University：Natural Science Edition

基　　金：国家重大基础研究发展计划(973计划)资助项目(2002CB312102);国家自然科学基金资助项目(60672094)

摘　　要：提出了一种基于隐马尔可夫模型和并行模型组合的特征补偿算法.首先,利用一个包含较多状态的隐马尔可夫模型来描述全部单词特征向量的分布.然后,根据静音段估计的噪声均值和方差,采用并行模型组合方法调整隐马尔可夫模型的均值向量和协方差矩阵,使之与识别环境相匹配.最后,根据基于状态转移矩阵压缩的前向后向算法计算隐马尔可夫模型的后验概率,并通过最小均方误差准则估计纯净语音特征向量.实验结果表明,该算法能够更加准确地估计纯净语音特征向量,其性能明显优于基于高斯混合模型的特征补偿算法;状态转移矩阵压缩算法可以在不影响补偿精度的前提下,显著减少前向后向算法的计算量.A feature compensation algorithm based on hidden Markov model （HMM） and parallel model combination （PMC） is presented. Firstly, a HMM composed of a number of states is employed to represent the distribution of the speech features of all words. Then, according to the mean and covariance of noise from noise-only frames, the mean vectors and covariance matrices of the HMM are transformed to the testing condition by the PMC method. Finally, the posterior probability of HMM is computed by the forward-backward algorithm based on the compression of the state transition matrix, and the clean speech feature is calculated by the minimum mean squared error method. The experimental results show that the proposed algorithm can restore the clean speech feature more accurately and outperforms the feature compensation algorithm based on Gaussian mixture model （GMM）. Besides, the state transition matrix compression method can greatly reduce the computational cost of the forward-backward algorithm without decreasing the compensation performance.

关 键 词：语音识别 特征补偿 隐马尔可夫模型 并行模型组合 

分 类 号：TN912.34[电子电信—通信与信息系统]