Recursive implementations of informed spatial filters
Informed spatial filters (ISFs) have been shown to provide high-quality speech acquisition in dynamic scenarios due to their ability to almost instantaneously adapt the filter coefficients based on the statistics of the desired and undesired signals. In most contributions, ISFs have been implemented in closed form as minimum variance distortionless response (MVDR), or minimum-mean-squared error filters. The goal in this paper is to discuss and evaluate recursive implementations of ISFs. We show that the implementations in a generalized sidelobe canceller (GSC) structure are not equivalent to the closed form MVDR, due to the fact that the filter coefficients of both implementations are updated at each time-frequency bin. The complexity of the implementations is discussed and experimental evaluation is performed for different dynamic scenarios where the goal is to extract a desired speaker in the presence of interfering speakers.