量化子空间分布隐马尔可夫模型的直接训练

Direct Training of Quantized Subspace Distribution Hidden Markov Model

尽管量化子空间分布隐马尔可夫模型 ( QSDHMM)的间接训练算法具有简单实用等优点 ,但仍存在两个方面的不足 :其一 ,QSDHMM的间接训练实际上要经历两个最优化过程 ,即先用原始语音数据训练连续分布隐马尔可夫模型 ( CDHMM) ,然后将训练好了的 CDHMM转换成 QS-DHMM.因此 ,QSDHMM的精度将受到影响 ;其二 ,没有发挥 QSDHMM本身参数少的潜在优势 .在系统的子空间捆绑结构为已知的前提下 ,文中提出了 QSDHMM的直接训练算法 .仿真实验表明 ,与传统的训练算法相比较 ,采用直接训练算法可减少训练数据约 1 0倍 。

Though the indirect training algorithm of the quantized subspace distribution hidden Markov model (QSDHMM) is simple and applicative. It has two drawbacks: (1) It involves two separate optimization processes——first training continuous distribution hidden Markov model (CDHMM), then converts to QSDHMM, so the accuracy of QSDHMM is not guaranteed to be optimal. (2) It does not take potential advantage of the fewer model parameters in QSDHMM. This paper devises a QSDHMM training algorithm to estimate QSDHMM parameters directly from speech data if having a prior knowledge of the subspace distribution tying structure of the system (SDTS). Computer simulation shows that using direct training algorithm to estimate QSDHMM parameters can reduce the training data roughly 10 times in comparison with the traditional algorithm, but almost no loss in recognition accuracy.