摘要
针对传统子空间语音增强方法中,由于不能去除整个噪声子空间而导致语音特征值估计的偏差,致使增强语音中仍有残留噪声的问题,提出一种新的用小波包改进的方法,该方法利用小波包对噪声的抑制功能,首先对带噪语音进行KL(Karhunen-Loeve Transform)变换,得到带噪语音的特征值,并对该特征值进行Daubechies8小波尺度分解,利用新的改进的软判决阈值函数去除一部分噪声子空间;然后再在子空间内用统计信息的方法实时跟踪此时噪声特征值,进一步消除所有噪声子空间,从而得到最终估计的语音特征值;最后由KL逆变换还原出纯净语音。仿真结果表明,在输入信噪比相同的情况下,经过该方法的增强语音的输出信噪比明显高于传统子空间方法,听觉感受上增强语音也具有更好的清晰度和可懂度。
In the traditional subspace speech enhancement method, the noise subspace can not be reduced that leads to the estimating bias of speech eigenvalues, and results in residual noise in enhanced speech. A new improved method using wavelet packet was advanced,which is based on the noise suppression of wavelet packet. Karhunen-Loeve Transform to the nosiy speech was con- ducted firstly, its eigenvalues were attained and then decomposed by db8 wavelet packet, a part of noise subspace was reduced by new modified wavelet soft-decision threshold function. Second- ly,in the subspace residual noise eigenvalues were estimated by statistical information method to further remove the noise subspace, and then attain final estimated speech eigenvalues. Lastly,the clean speech was restored by KL inverse transform. The simulation results show that the new method has higher SNR in the output side, the better speech quality than the traditional method in the same input SNR condition.
出处
《太原理工大学学报》
CAS
北大核心
2011年第2期117-120,共4页
Journal of Taiyuan University of Technology
基金
国家自然科学基金资助项目(61072087)
山西省自然科学基金资助项目(2010011020-1)
关键词
语音增强
子空间
小波包
阈值函数
speech enhancement
subspace
wavelet packet
threshold function
