卷积混合语音的联合块对角化盲源分离算法

BLIND SOURCE SEPARATION ALGORITHM FOR CONVOLUTIVE SPEECH MIXTURES USING JOINT BLOCK-DIAGONALIZATION

针对语音卷积混合模型,提出了一种新的时域盲源分离算法。首先对观测信号进行重新排列,将卷积混合盲分离问题转化为瞬时混合盲分离问题,然后对联合近似对角化算法进行了推广,利用语音的非平稳和短时平稳特征定义联合差分相关矩阵和联合块对角化代价函数,通过鲁棒的白化过程和求解最优化问题实现卷积语音的盲分离。由于避免了时域卷积运算和变换域处理,使算法更加简单,复杂度更低。仿真结果验证了该算法的有效性,同时,就数据长度参数变化对信干比的影响,以及通过与基于线性预测的卷积盲分离算法和自然梯度卷积盲分离算法的比较对该算法的性能做了进一步的分析。

The method proposed resets the sampling mixture signals and transforms the convolutive BSS into instantaneous BSS. The joint approximate diagonalization algorithm is generalized, making use of the non-stationarity and short-time stationarity of speech sources, and the joint difference correlation matrix and joint block-diagonalization cost function are defined. Through robust whitening process and resolving optimization problem, the covolutive blind speech source separaton is realized. The algorithm avoids the convolution calculation and domain transformation to decrease the complexity. Computer simulation shows the effectiveness, while the further performance analysis is done by comparing the linear-pridiction-based convolutive BSS algorithm with the natural gradient convolutive BSS algorithm. The effect of data length parameter on the signal to interference ratio(SIR) is also discussed.