一种用于语音增强的频域盲信号分离算法

研究了在未知声源信息和传声器空间位置的情况下,利用盲信号分离的方法实现语音增强。通过把基于信息论的信息最大化算法推广到频域,使得时域的卷积混合问题转变为频域的瞬时混合问题,进而就可以在每个频段分别进行独立分量分析,分离效果有明显改进,算法收敛性也得到提高。为了克服在频域中实现盲分离时所固有的位序不确定性和比例缩放问题对分离性能的严重影响,采用聚类的方法对每个频率段的分离结果进行排序。对真实环境中录制的语音、音乐混合信号和语音、语音混合信号进行了计算机仿真,分离之后使语音的信噪比提高了10-15dB,很好地实现了语音增强的目的。

A frequency domain blind source separation algorithm for speech enhancement

In order to enhance speech signals acquired with microphone arrays in a noisy environment, without a priori information about the sources and the placement of the microphones, blind source separation (BSS) based on information maximization is proposed and then extended to the frequency domain. In this way the convolutive problem can be inverted to an instantaneous one, with independent component analysis (ICA) performed separately in every frequency bin. Efficiency of the separation and convergence can be improved through this transformation. However, the frequency domain BSS has the inherent problem of permutation and scaling, which can severely affect the performance. According to this, a clustering method is employed in every frequency bin. Finally, satisfactory experimental results are obtained in computer simulation with data recorded in a real meeting room. The SNR improvement can reach 10-15dB after blind source separation, and the speech quality is remarkably improved.