文章摘要
王振力,刘志华,白志强.基于卷积盲源分离的噪声鲁棒性语音识别的研究[J].声学技术,2009,(3):276~279
基于卷积盲源分离的噪声鲁棒性语音识别的研究
A study of noise robust speech recognition based on convoluted blind source separation
投稿时间:2008-04-07  修订日期:2008-06-20
DOI:
中文关键词: 噪声鲁棒语音识别  盲信号分离  MFCC  解相关
英文关键词: noise robust speech recognition  blind source separation  Mel-Frequency Cepstral Coefficients  decorrelation
基金项目:江苏省博士后科研基金资助项目(0701008C);中国博士后科学基金(20070420561)
作者单位
王振力 南京国际关系学院, 南京, 210039 
刘志华 南京国际关系学院, 南京, 210039 
白志强 南京国际关系学院, 南京, 210039 
摘要点击次数: 938
全文下载次数: 1280
中文摘要:
      研究了一种基于卷积盲分离算法与MFCC(Mel-Frequency Cepstral Coefficient)特征相结合的噪声鲁棒语音识别方法。该方法在预处理阶段,首先计算预白化观测数据的多阶自相关协方差矩阵,以获得多时延处理的二阶解相关统计信息。然后利用得到的二阶统计信息构建两个对称正定矩阵,通过Cholesky因式分解等一系列变换获得唯一存在的矩阵,根据此矩阵估算语音信号并提取MFCC特征用于后续识别。实验结果表明,在低信噪比条件下,该方法对于数字语音的识别性能优于基本的MFCC识别器和文献中已有的卷积分离算法。
英文摘要:
      A method of noise robust speech recognition is investigated by combining a blind source separation algorithm and MFCC(Mel-Frequency Cepstral Coefficients).At the pretreatment phase of speech recognition,this method obtains the multi-lags decorrelated second-order statistics information by computing the prewhitened covariance matrix and its multi-sample delayed counterpart.The obtained second-order statistics is used for forming two positive definite symmetry matrices.A unitary matrix is achieved by applying a set of transforms such as Cholesky decomposition and SVD(singular value decomposition) to the two formed matrices.The speech signal is estimated according to this unitary matrix and MFCC is then computed for the following speech recognition.Experimental results indicate that this method performs better than the unitary MFCC and the convoluted separation algorithm in literature for low signal-to-noise ratio(SNR).
查看全文   查看/发表评论  下载PDF阅读器
关闭