基于状态驻留时间的汉语语音分段概率模型

为了解决分段概率模型 (SPM)因缺少对时间信息描述而带来的建模精度低的问题 ,提出了状态驻留分段概率模型 (SDSPM)。SDSPM中包含了用伽玛分布表示的状态驻留概率 ,以刻划语音的时间特征。此驻留概率相当于隐马尔可夫模型 (HMM)中的状态转移概率 ,但使 SDSPM描述语音时间特征的能力强于 HMM。SDSPM既改善了 SPM的模型性能 ,同时又避免了 HMM的计算复杂度问题。测试实验证明了 SDSPM模型在汉语语音识别中的有效性。

State duration-based segmental probability model for Chinese speech

This paper presents the State Duration based Segmental Probability Model (SDSPM) to cope with low modeling precision during the analysis of Chinese speech, which is basically caused by the lack of proper time information during the speech processing. The probability of state duration, represented by a Gamma distribution, is used to depict the time characteristics of the speech. The analysis shows that the duration probability is analogous to the state transition probability, but is more suitable for representing the time information. The SDSPM performance is better than the SPM performance, while avoiding the problem of the high computational complexity of HMM. The SDSPM feasibility is tested in an actual Chinese speech recognition environment.