基于区域增长校正的频域盲源分离排序算法

卷积盲源分离可以在频域得到有效解决，但频域盲源分离必须解决排序模糊问题。该文提出一种基于区域增长校正的频域盲源分离排序算法。首先对卷积混合信号短时傅里叶变换，在频域的各个频点处建立瞬时模型进行独立分量分析，在此基础上使用分离信号功率比的相关性，对所有频点进行逐点排序置换。其次根据阈值将排序后的结果划分为若干个小区域。最后按区域增长方式进行区域置换与合并，最终得到正确的分离信号。区域增长校正可最大限度地减少频点排序错误扩散现象，从而改善分离效果。在模拟和真实环境中分别进行语音盲源分离实验，结果表明所提算法的有效性。

Frequency Domain Blind Source Separation Permutation Algorithm Based on Regional Growth Correction

The convolutive blind source separation can be effectively solved in frequency domain, but blind source separation in frequency domain must solve the problem of ranking ambiguity. A frequency-domain blind source separation sorting algorithm is proposed based on regional growth correction. First, the convolutional mixed signal short-time Fourier transform is used to establish an instantaneous model at each frequency point in the frequency domain for independent component analysis. Based on this, the correlation of the power ratio of the separated signal is used to sort all frequency points one by one replacement. Second, according to the threshold, the sorted result is divided into several small areas. Finally. regional replacement and merging is performed according to the regional growth method, and the correct separation signal is finally obtained. Regional growth correction minimizes the mis-proliferation of frequency sorting and improves separation results. The speech blind source separation experiments are performed in the simulated and real environments respectively. The results show the effectiveness of the proposed algorithm.