A novel fuzzy clustering algorithm using observation weighting and context information for reverberant blind speech separation

Time-frequency masking has evolved as a powerful tool for tackling blind source separation problems. In previous work, mask estimation was performed with the help of well-known standard cluster algorithms. Spatial observation vectors, extracted from a set of microphones, were grouped into separate clusters, each representing a particular source. However, most off-the-shelf clustering methods are not very robust to outliers or noise in the data. This lack of robustness often leads to incorrect localization and partitioning results, particularly for reverberant speech mixtures. To address this issue, we investigate the use of observation weights and context information as means to improve the clustering performance under reverberant conditions. While the observation weights improve the localization accuracy by ignoring noisy observations, context information smoothes the cluster membership levels by exploiting the highly structured nature of speech signals in the time-frequency domain. In a number of experiments, we demonstrate the superiority of the proposed method over conventional fuzzy clustering, both in terms of localization accuracy as well as speech separation performance.