基于频域ICA的语音特征增强  

Speech feature enhancement based on frequency-domain ICA

作　　者：吕钊[1,2] 吴小培[1] 李密[2] 

机构地区：[1]安徽大学计算智能与信号处理教育部重点实验室,合肥230039 [2]空军第一航空学院航空电子工程系,信阳464000

出　　处：《振动与冲击》2011年第2期238-242,257,共6页Journal of Vibration and Shock

基　　金：国家自然科学基金资助项目(60771033);博士点基金(200803570002)

摘　　要：为了降低卷积噪声对语音特征所产生的影响,提高语音识别正确率,在此提出了一种基于频域ICA(Independent Component Analysis,独立分量分析)的语音特征增强算法。该算法首先使用频域ICA方法作对噪声进行估计,然后在倒谱域内将带噪语音信号的短时谱减去所估计噪声的短时谱,最后根据去噪后语音信号的短时谱计算美尔倒谱系数(MFCC)作为特征参数。在仿真和真实环境下的语音识别实验中,所提出的语音特征参数相比较传统的MFCC其识别正确率分别提升了38.2%和35.8%。实验结果表明该算法能够较好地解决卷积噪声环境下训练与识别特征不匹配的问题,有效提高了语音识别系统的识别正确率。To suppress the interference of convolutive noise on speech features and improve the rate of speech recognition,a speech feature enhancement algorithm based on frequency-domain ICA（independent component analysis） was presented here.In the proposed algorithm,noise short-time spectrum was estimated using the frequency-domain ICA algorithm,and then noise reduction was achieved by subtracting the estimated noise short-time spectrum from a noisy speech short-time spectrum to be enhanced in the Mel-scale filter bank domain.As a result,the robust MFCC（Mel-frequency cepstral coefficients） were acquired.Simulation and real environment test results revealed that their recognition raties with the proposed algorithm increase about 38.2% and 35.8% respactively compared with that of the conventional MFCC;the mismatch between training features and identifing ones in convolutive noise environment can be suppressed effectively with the proposed algorithm.

关 键 词：频域ICA 语音 特征增强 美尔倒谱系数(MFCC) 

分 类 号：TN912.34[电子电信—通信与信息系统]