利用韵律信息的CHMM连续数字语音识别  

A Study of Connected Digit Speech Recognition Based CHMM with Prosodic Information

作　　者：张静亚[1] 俞一彪[2] 

机构地区：[1]常熟理工学院物理与电子科学系,江苏省常熟市215500 [2]苏州大学电子信息学院,江苏省苏州市215021

出　　处：《电子工程师》2006年第12期43-46,共4页Electronic Engineer

基　　金：江苏省高校自然科学基金重点项目(04KJA51033)

摘　　要：提出了一种结合韵律信息的高性能汉语连续数字语音识别算法,该识别算法基于CHMM(连续隐马尔可夫模型),采用MFCC(MEL频率倒谱系数)为主要语音特征参数,结合韵律信息进行连续数字精确分割,能够有效区分易混数字。算法采用两级识别框架来提高语音识别率,其中,第1级对连续数字分割,在此基础上进行数字语音识别,输出各候选结果,第2级在候选结果中确定易混数字对,并运用韵律信息进一步选择正确结果。实验表明,最终汉语连续数字语音识别率有很大提高。A new algorithm for connected digital speech recognition based on CHMM using prosodic information is proposed. Every digit is modeled by a five-state CHMM described by MFCC coefficients. With the prosodic information, the connected speech is separated precisely, and the digits which acoustic features used to confuse easily can be recognized correctly. The algorithm employs two-level scheme. In the first level, the input speech is separated into individual digital syllables, and then the syllables are recognized and will output the first two digital candidates with higher scores. In the second level, the right digit string is extracted from the candidate lattice using the prosodic information. Experiments show that the proposed algorithm can improve the connected digital speech recognition performance.

关 键 词：语音识别 连续隐马尔可夫模型(CHMM) 韵律信息 

分 类 号：TN912.34[电子电信—通信与信息系统]