惩罚概率对经典隐马尔可夫模型(HMM)齐次假设的补偿

Compensation for the homogeneous assumption of classical HMM by probability penalty

分析了经典隐马尔可夫模型（Ｈｉｄｄｅｎ Ｍａｒｋｏｖ Ｍｏｄｅｌ，ＨＭＭ）齐次假设的理论缺陷，以及两种非齐次ＨＭＭ．语音识别对比实验表明，惩罚概率法是稳健的、且更有效的补偿方法．

The theoretical weakness of classical Hidden Markov Model (HMM), i.e. its homogeneous assumption and two inhomogeneous HMMs are studied in this paper. Control experbent for speeCh recognition shows that the penalty approach is robust and more effective in compensating this unreasonable assumption.