一种对应力变异语音的特征补偿方法  

A compensation method for recognition of speech under G-force

作　　者：张磊[1] 韩纪庆[1] 王承发[1] 

机构地区：[1]哈尔滨工业大学计算机科学与工程系

出　　处：《声学学报》2004年第1期18-22,共5页Acta Acustica

基　　金：国家自然科学基金资助项目(60085001)

摘　　要：提出一种描述正常语音和变异语音之间关系的补偿因子。该补偿因子兼顾考虑了由于变异引起的特征分布中均值和方差的变化,并在k-均值初始化的参数基础上,采用期望最大化(EM)算法迭代估计变异补偿因子的值。通过估计出的补偿因子对变异语音特征进行补偿。对航空模拟飞行器中采集的应力变异下特定话者小词表孤立词的实验结果表明,利用所提出的方法可以将识别率提高32.3％。A compensation factor is proposed, which can describe the relations between the neutral speech and the speech under stress. And the stress compensation factor also considers both the shift of means and the differences of the variances caused by the stress. Based on the parameters initialized by k-means method, the final results of the compensation factor are estimated by the EM algorithm. Through the compensation factor, the performance of a recognition system can be enhanced. Under the G-force stress, the proposed method can get 32.3% improvement for a small-vocabulary speaker-dependent system.

关 键 词：应力变异语音 特征补偿 正常语音 迭代估计 航空模拟 识别率 语音识别系统 

分 类 号：TN912.3[电子电信—通信与信息系统]