基于波束形成与最大似然估计的近距离双麦克风语音增强算法

Beamforming and maximum likelihood estimation for speech enhancement using dual closely-spaced microphones

为了解决波束形成法直接应用于近距离双麦克风系统时存在的问题,如目标信号低频段能量损失、多方向的竞争性语音噪声难以被有效抑制等,该文提出一种基于波束形成与最大似然估计的两步去噪方法。该方法首先使用加权叠加滤波器对混合声源进行时频分解,然后通过2个零点分别在0°和180°的波束形成图的幅频响应比值,设置各时频单元所对应的初步掩蔽值,在避免低频滚降现象出现的情况下,抑制本底噪声;最后根据统计模型和简化的最大似然估计法,抑制多方向的竞争性语音噪声,进一步增强目标信号。测试结果表明：在低信噪比、多种类型噪声源共同存在的情境下,该方法可以在无需低通滤波或宽带波束补偿的情况下,恢复原始信号的能量分布特点,明显提升信噪比。

Traditional beamforming systems using dual closely-spaced microphones have various problems such as low-frequency-roll-off and limitations in suppressing competitive speech noises from multiple directions. This paper presents a two-step beamforming and maximum likelihood estimation algorithm. The algorithm first uses a WOLA filter for the time-frequency analysis for the speech mixture and then sets mask values to suppress background noise without low-frequency-roll-off based on the ratio of the two beamforming patterns, which have zeros at 0°and 180°. A statistical model and the maximum likelihood estimation are then used to further enhance thespeech. Tests indicate that the algorithm effectively recovers the energy distribution of the target signal and improves the signal-to-noise ratio without a low-pass filter or broadband compensation when the signal-to-ratio is low or multiple kinds of noises exist.