Dereverberation with an Iterative Least-Squares Technique and Minimum Mean-Square Error Estimation for Automatic Speech Recognition

This work is about dereverberation for automatic speech recognition. The use of a linear minimum mean-square error estimator for enhancing a recently proposed dereverberation method is investigated. The conducted phoneme recognition experiments show that the resynthesis step, which was done in the original work of the dereverberation method, can be omitted. Furthermore, it is shown that the recognition performance can be increased with the proposed estimator approach under certain reverberant conditions.