Beamforming Initialization and Data Prewhitening in Natural Gradient Convolutive Blind Source Separation of Speech Mixtures

Successful speech enhancement by convolutive blind source separation (BSS) techniques requires careful design of all aspects of the chosen separation method. The conventional strategy for system initialization in both time- and frequency-domain BSS involves a diagonal center-spike FIR filter matrix and no data preprocessing; however, this strategy may not be the best for any chosen separation algorithm. In this paper, we experimentally evaluate two different approaches for potentially-improving the performance of time-domain and frequency-domain natural gradient speech separation algorithms - prewhitening of the signal mixtures, and delay-and-sum beamforming initialization for the separation system - to determine which of the two classes of algorithms benefit most from them. Our results indicate that frequency-domain-based natural gradient BSS methods generally need geometric information about the system to obtain any reasonable separation quality. For time-domain natural gradient separation algorithms, either beamforming initialization or prewhitening improves separation performance, particularly for larger-scale problems involving three or more sources and sensors.