摘要
频域的语音信号盲源分离多采用短时傅里叶变换以及Wigner-Ville分布(WVD)求信号的功率谱,而短时傅里叶变换对于多分量信号的频率分辨率受窗函数影响很大,WVD是一种非线性时频变换,处理多分量信号受交叉项影响很大。局部多项式傅里叶变换(LPFT)不仅提高了频率估计精度而且大大减少了时频分布中交叉项的影响。将语音信号表示为多分量的多项式相位信号,对语音信号作二阶LPFT,求得其局部多项式傅里叶变换谱(LPP),并构造时频矩阵,采用联合近似对角化算法求得能使信号功率谱矩阵近似对角化的一个酉矩阵,通过信号的白化以及酉矩阵来估计源信号,有效地分离出了原始信号。仿真结果表明,在噪声环境下可以将两个不同的语音信号进行分离。
The speech signal blind source separation in frequency domain uses short-time Fourier transformation （STF＇F） and Wigner-Ville distribution （WVD）for signal power spectrum. For the multi component signal short-time Fourier transform frequency resolution is affected by window function. WVD is a non-linear time-frequency transform, in the processing of the multi-component signal is highly affected by the cross terms. Local Polynomial Fourier Transform （LPF＇F） has been used in many signal processing field as a generalized form of STFI＇. The precision of frequency estimation is improved and the influ- ence of cross terms in the time-frequency distribution is reduced. The speech signal is expressed by the multi-component pol- ynomial phase signal in this paper, and makes the second-order LPF＇F. Its local polynomial Fourier conversion spectrum （LPP） is obtained and the time-frequency matrix is coonstructed. Finally the joint approximate diagonalization algorithm is used to obtain the enabling signal＇ s power frequency spectrum matrix that approximate diagonalization unitary matrix. The source signal through the signal albinism as well is estimated as unitary matrix, it can separate the primary signal effectively. The results show that the different speech signals can be successfully separated in the noise environment.
出处
《电声技术》
2014年第2期45-49,共5页
Audio Engineering
基金
辽宁省教育厅基金项目(L2010420)
