Enhanced Online IVA with Switched Source Prior for Speech Separation

Independent vector analysis (IVA) is a blind source separation (BSS) technique that has demonstrated efficiency in separating speech signals from their convolutive mixtures in the frequency domain. Particularly, it avoids the problematic permutation problem by using a multivariate source prior to model statistical inter dependency across the frequency bins of each source signal. The selection of the source prior is vital to the performance of the method. Practical real time systems require an online mode which is performed iteratively as signal data arrive. The performance of the online IVA is measured by the convergence time and steady state separation and accuracy. This paper proposes a novel switched source prior technique to improve the performance of the online IVA algorithm. The techniques switches between two source priors to acquire the better performance properties of both distributions at different stages of the learning algorithm. The switching process is controlled by an adaptive learning scheme as a function of proximity to the target solution. The experimental results demonstrate an enhanced separation performance using real room impulse responses and recorded speech signals.