相干函数在双传声器语音增强中被广泛用于抑制非相干噪声,但其对部分相干的扩散噪声和强相干的方向性干扰缺乏处理能力,经常需要额外算法步骤另行处理。本文提出在互谱估计的的过程中,引入对目标信号方向的先验,当单个互谱样点的相位处在目标区域之外时,将其相位随机化,然后进行时间平均,得到一个修改的互谱估计,进而可以得到修改的相干函数估计。其可以在基于双传声器的语音增强任务中,直接作为增益函数应用于幅度谱修正框架中,达到同时抑制非相干噪声和方向性相干干扰的目的。统计仿真实验和实际实验均验证了本文方法的有效性。
Coherence function is widely used in two microphone speech enhancement to suppress non - coherent noise, but it cannot cope with partially coherent noise and highly coherent directional interference which must be handled by additional algorithms. This pa- per proposes to add the a priori on the direction of the target signal to the procedure of cross spectrum estimation by randomizing the phase of the cross spectrum sample when its phase is not in the pre - specified region. After averaging along the time, a modified cross spectrum and subsequently a modified coherence function estimator can be obtained. It can be applied as the gain function in the spec- tral modification framework for two microphone speech enhancement to suppress non - coherent noise and directional coherent signal simultaneously. Statistical simulation and real environment experiment verifies the effectiveness of the proposed method.