Speech dereverberation and speaker separation using microphone arrays in realistic environments

Abstract

This thesis concentrates on comparing novel and existing dereverberation and speaker separation techniques using multiple corpora, including a new corpus collected using a microphone array. Many corpora currently used for these techniques are recorded using head-mounted microphones in anechoic chambers. This novel corpus contains recordings with noise and reverberation made in office and workshop environments. Novel algorithms present a different way of approximating the reverberation, producing results that are competitive with existing algorithms. Dereverberation is evaluated using seven correlation-based algorithms and applied to two different corpora. Three of these are novel algorithms (Hs NTF, Cauchy WPE and Cauchy MIMO WPE). Both non-learning and learning algorithms are tested, with the learning algorithms performing better. For single and multi-channel speaker separation, unsupervised non-negative matrix factorization (NMF) algorithms are compared using three cost functions combined with sparsity, convolution and direction of arrival. The results show that the choice of cost function is important for improving the separation result. Furthermore, six different supervised deep learning algorithms are applied to single channel speaker separation. Historic information improves the result. When comparing NMF to deep learning, NMF is able to converge faster to a solution and provides a better result for the corpora used in this thesis.