On Joint Dereverberation and Source Separation with Geometrical Constraints and Iterative Source Steering

In order to improve both the separation performance and the convergence speed, several geometrically constrained independent vector analysis (GC-IVA) algorithms have been developed. Those algorithms are based on the multiplicative transfer function model, which assumes that the analysis window length is longer than the effective part of the room impulse responses. However, this assumption does often not hold in reverberant environments, particularly if the reverberation is strong, which makes the algorithms suffer from significant performance degradation. To circumvent this issue, an algorithm was developed, which jointly optimizes the weighted prediction error (WPE) dereverberation method and GC-IVA (GC-WPE-IVA). While it has demonstrated promising performance, this joint optimization method involves matrix inversion; so it is computationally very expensive. This work attempts to improve the efficiency and stability of GC-WPE-IVA. We develop an iterative source steering (ISS) updating algorithm in the framework of GC-WPE-IVA. The experimental results show that the developed method is computationally much more efficient yet it can achieve comparable separation performance in reverberation environments as compared to GC-WPE-IVA.