MFCC特征加权应力影响下的变异语音识别  

Recognition of speech under G - force based on weighted features of MFCC

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技术学院,黑龙江哈尔滨150001

出　　处：《哈尔滨工业大学学报》2002年第6期743-747,共5页Journal of Harbin Institute of Technology

基　　金：国家自然科学基金资助项目(60085001);教育部留学人员科研启动基金资助项目.

摘　　要：为了提高语音识别系统的顽健性,在对应力影响下语音数据分析的基础上,提出一种新的基于MF-CC系数加权的变异语音识别方法.它首先通过正常语音和应力影响下变异语音的差异求得一个变异影响因子,然后对该因子的倒数进行规正作为MFCC特征不同维的权值,从而减弱受变异影响较大的特征对识别性能的影响.对航空模拟飞行器中采集的特定话者小词表孤立词的实验表明,与传统的多重风格训练方法相比,该方法的识别率提高了10.9％;将其和倒谱平均减方法相结合,可进一步将识别率提高5.4％.For the robustness of speech recognition system, based on the analysis of speech under G - force, a new method of weighted features of MFCC is proposed for the stressed speech recognition, in which a stress effect factor is firstly obtained by the difference between neutral speech and speech under G - force, and then the weights are gotten by using the inverse of the stress effect faetor, and thus the effect of the most varied features can be weakened. For speaker dependent, isolated words selected from an aero -flight simulator, using the proposed method, an improvement of 10. 9% is achieved relative to the traditional multi - style method. When the proposed method is combined with the CMS, a further improvement of 5.4% is gotten.

关 键 词：MFCC 特征加权 语音识别 应力 变异语音 MEL频率倒谱系数 

分 类 号：TN912.34[电子电信—通信与信息系统]