Multi-microphone speech dereverberation using LIME and least squares filtering

In this paper a multi-microphone speech dereverberation algorithm is presented. The developed algorithm is based on the LIME algorithm proposed by Delcroix et al. in [IEEE Trans. Audio, Speech and Language Process., vol. 15,no. 2, pp. 430-440, Feb. 2007]. The LIME algorithm is shown to be signal dependent and to either produce a very good or a very poor estimate of the source signal. Two non-intrusive methods are proposed to assess the performance of the LIME algorithm for an unknown source signal. These methods can be used to detect errors in the source signals' estimation. Least squares filters are calculated using signal segments that are successfully recovered by LIME, and are used to dereverberate other signal segments for which LIME produced poor estimates. Experimental results demonstrate that the signal segments for which LIME fails can successfully be detected and dereverberated using the least squares filters.