【摘要】 方言的差异性在语音层面上反映在时间序列结构的不同。传统的语音建模方法只能建立稳定的时间序列结构,而方言语音是典型的动态时变时间序列结构。为了更好地提取方言时间序列结构,文中采用动态贝叶斯网路(DBN)进行建模分析,并对DBN的构建方法进行了研究,这种结构与常用于语音识别中的隐马尔可夫模型的不同之处在于它揭示多个时间片内的节点之间的影响。文中探索了不同结构和参数对识别效果的影响。文中的研究表明动态贝叶斯网络对汉语方言的识别比传统方法要好,识别率达到了98.9%。更多还原

【Abstract】 The differentiation of Chinese dialect is the different time series in the phonetic.Traditional speech modeling methods can only establish time series,but the dialect speech is typical time-varying series.It chose dynamic Bayesian networks to model the speech in order to extract the time series structure of dialect speech.It also studied the method to model the DBN structure and the influence of the model complexity on recognition rate.The structures of this paper is more complex than the HMM because these structures notice the influence of the nodes in more than two time series.Experiments show that the DBN method is an excellent method with high rate 98.9%.更多还原