一种强混响环境下的盲语音分离算法

摘要

In strong reverberant environment, blind separation of speech signals under convolutive mixture is a difficult problem,since the conventional algorithms will degrade quickly due to long reverberant time., which is longer than the nonstationary time of speech. A new method is proposed to deal with this case, where two short time Fourier transforms (STFT) are used to transform time signals into frequency domain, so the convolutive mixture with long reverberant time in time domain can be expressed as the instantaneous mixture in the modulation spectrogram, and some algorithms for instantaneous mixture can be applied easily. In this paper, the Independent Vector Analysis (IVA) algorithm is proposed to avoid permutation ambiguity efficiently in the modulation spectrogram. The computer simulations confirm that the proposed algorithm outperforms conventional frequency domain BSS algorithms in strong reverberant environment.%强混响环境下语音信号的频域盲分离问题是盲源分离领域的一个难点,主要是因为混合系统的脉冲响应时间过长,甚至超过信号的非平稳时间,导致算法性能下降.本文针对这个问题提出了一种解决方法,在用一个短时傅立叶变换将时域卷积混合信号转化为频域后,再在时频域的不同帧间使用另一个短时傅立叶变换,将信号变换到调制谱域.时域上由较长的脉冲响应得到的卷积混合就被转化为调制谱域上的瞬时混合形式,而瞬时混合情形则采用独立向量分析(ⅣA)算法来避免排序模糊性问题.计算机仿真实验证实了该算法在强混响环境下优于传统频域盲分离算法.