语音信号频域盲分离中一种频率对准方法

Solution to Permutation Problem in Frequency-domain Blind Audio Separation

混合语音信号可以使用盲分离频域解法,对观测信号在每一个频点分别进行复值独立分量分析(CICA)算法来解混并得到分离信号,但带来了幅值和次序不定问题(后者又称频率对准)。讨论了频率对准算法中基于DOA估计的方法,并提出了一种基于分离矩阵初始化的频率对准方法,此方法易于实现。通过仿真表明,该方法较好地解决了次序不定问题,对卷积混合语音信号有较好的分离效果。

Convolved audio signals can be separated by frequency-domain blind source separation. Complex independent component analysis is performed in each frequency bin of the mixed signal, but the permutation inconsistency and scaling arbitrary are involved. The permutation alignment methods using DOA estimation is discussed and a new scheme based on separation matrix initialization is proposed. The simulation shows that the proposed method is easy to implement and can resolve the permutation problem. A good separation result is achieved.