基于改进的隐马尔科夫模型的语音识别方法

A speech recognition method based on improved hidden Markov model

针对隐马尔可夫(HMM)语音识别模型状态输出独立同分布等与语音实际特性不够协调的假设以及在使用段长信息时存在的缺陷,对隐马尔可夫模型进行改进,提出马尔可夫族模型。马尔可夫族模型可看作一个数学上由多个马尔可夫链构成的多重随机过程,HMM模型则是双重随机过程,因而,HMM模型可视为马尔可夫族模型的特例。马尔可夫族模型用条件独立性假设取代了HMM模型的独立性假设。相对条件独立性假设,独立性假设是过强假设,因而,基于马尔可夫族模型的语音模型更符合语音实际物理过程。在马尔可夫族语音识别模型中引入状态段长信息,能自动根据语速对语音单元段长进行调整。非特定人连续语音实验结果表明,利用状态段长信息的改进语音识别模型比经典HMM模型的性能明显提高。

In order to overcome the defects of the duration modeling of homogeneous hidden Markov model （HMM） in speech recognition and the unrealistic assumption that successive observations are independent and identically distribution within a state, Markov family model （MFM） was proposed. In the speech recognition model based on HMM, the time-sequence structure of speech signal was considered to be a double stochastic process, while Markov family model was a multiple stochastic process which consists of a few Markov chains, so HMM could be considered to be a special case of MFM. Moreover, independence assumption in HMM was placed by conditional independence assumption in MFM, and from the view of the statistics, the assumption of independence is stronger than that of conditional independence, so speech recognition model based on MFM is more realistic than HMM recognition mode. Markov Family model was applied to speech recognition, and duration distribution based MFM recognition mode which takes duration distribution into account and integrates the frame and segment based acoustic modeling techniques, was proposed. The speaker independent continuous speech recognition experiments show that this new recognition model has better performance than standard HMM recognition models.