摘要
为进一步抑制噪声,提出一种基于特征值合并的语音增强算法。在经典的内嵌式预白化子空间方法的基础上,用特征值合并来提高语音质量。研究发现,对含噪语音的协方差矩阵进行特征值分解后,大特征值分量主要包含语音信息,而小特征值分量主要包含噪声,特征值分量按特征值从小到大排序后,剔除相邻的小特征值分量,可有效抑制噪声,提高语音质量。相比于其它方法,基于特征值合并的语音增强算法能有效工作于各种噪声环境中,显著提高信噪比,并有更好的语音可懂度。
In order to further suppress noise, a kind of speech enhancement algorithm based on Eigen-value merge was proposed. Eigen-values combination was used to improve the speech quality on the basis of the classic embedded pre-whitening subspace methods. The study shows that, after decomposing the covariance matrix of speech signals with noise, the larger Eigen-value component mainly includes speech information, and the smaller Eigen-value component mainly contains noise. Sorted by Eigen-values from small to big, the adjacent large Eigen-value component replaces with small Eigen-value component, which can effectively suppress noise and improve the quality of speech. Compared with other speech enhancement algorithm, this algorithm based on Eigen value merger can work effectively in a variety of noisy environment, significantly improves the SNR, and has better speech intelligibility.
出处
《系统仿真学报》
CAS
CSCD
北大核心
2016年第7期1622-1627,共6页
Journal of System Simulation
基金
国家自然科学基金(61272315
60842009)
浙江省自然科学基金(Y1110342)
关键词
语音增强
子空间方法
特征值分解
语音质量
speech enhancement
subspace method
eigen-value decomposing
speech quality
