隐Markov模型中状态停留时间的模型化

在用传统的ＨＭＭ（ＴＨＭＭ）刻画现实中的语音时有一个显然的缺点，那便是它不能合适地表征语音信号的时域结构。本文采用依赖于时间的状态转移概率来模型化状态停留时间，修改后的模型称为ＭＨＭＭ。对于参加过训练的说话人，ＴＨＭＭ和ＭＨＭＭ的正识率基本上差不多。而对于未参加过训练的说话人，ＭＨＭＭ的正识率明显高于ＴＨＭＭ的正识率。也就是说，ＭＨＭＭ对于说话人的适应性要好于ＴＨＭＭ。原因在于ＭＨＭＭ更多地包含了发音时音素间的跳转信息。

Modeling state duration in HMM

To characterize the nature of real speech using the traditional HMM scheme there is an obvious disadvantage, that is the traditional HMM does not represent the temporal structure of speech appropriately. In this paper, time dependent state transition probability was used to model state duration. We gained almost. The same recognition performances were gained by both HMM for trained speaker in multi speaker mode. However, it is found that the modified HMM is more robust than the traditional HMM for untrained speaker in multi speaker mode. It is due to the modified HMM contains more phonetic transition information about the syllable to be recognized.