一种新型汉字音节整体向量模型的识别研究  

Research on the recognition of the novel Chinese syllable model based on whole vectors

作　　者：贺苏宁[1] 虞厥邦[1] 

机构地区：[1]电子科技大学电子工程学院,四川成都610054

出　　处：《系统工程与电子技术》2005年第2期343-348,356,共7页Systems Engineering and Electronics

摘　　要：通过对hiddenmarkovmodel(HMM)和segmentmodel(SM)模型的简要分析,指出了它们的某些缺陷,提出了一种新的基于汉字音节整体的Melfrequencycepstrumcoefficients(MFCC)向量模型。该模型能够根据各个音节的持续时间动态地调整帧长,进而比较完整地表现了语音时频信息的演化过程。实验数据显示,在同样的测试条件下,对于上下文相同的同性语音,帧数固定比帧长固定的识别率改善3.0%以上。还分析了几个主要影响汉字语音识别率的参数。研究表明,参数设置是否得当对于识别率有一定的影响。By analyzing hidden Markov model (HMM) and segment model (SM) briefly, some deficiencies of these models are addressed, and a novel Chinese syllable model based on whole Mel frequency cepstrum coefficient (MFCC) vectors is proposed. The model can adjust dynamically the frame length by syllable duration, which fully presents the evolving process of time-frequency information of speech. It is show by experiments that in the same test condition, the recognition rate of fixed number of frames is 3.0% more than that of fixed length of frames. Some main parameters that affect the recognition rate are also analyzed. Whether these parameters are set properly or not will affect the recognition rate to a certain degree.

关 键 词：动态时间伸缩 隐马尔柯夫模型 分段模型 观测向量 时频特征分析 

分 类 号：TN911.21[电子电信—通信与信息系统]