[Paper Review] Spatially Informed Independent Vector Analysis

Andreas Brendel, Thomas Haubner+1 more|arXiv (Cornell University)|Jul 23, 2019

Blind Source Separation Techniques23 references4 citations

TL;DR

This paper proposes a Maximum A Posteriori (MAP) formulation of Independent Vector Analysis (IVA) that incorporates spatial priors via a directional constraint on demixing filters, enabling automatic resolution of the outer permutation ambiguity in blind source separation. Using a Majorize-Minimize (MM) algorithm, the method achieves convergence speed comparable to the state-of-the-art auxIVA while outperforming gradient-based spatially constrained IVA in interference suppression and computational efficiency across real-world room impulse responses.

ABSTRACT

We present a Maximum A Posteriori (MAP) derivation of the Independent Vector Analysis (IVA) algorithm, a blind source separation algorithm, by incorporating a prior over the demixing matrices, relying on a free-field model. In this way, the outer permutation ambiguity of IVA is avoided. The resulting MAP optimization problem is solved by deriving majorize-minimize update rules to achieve convergence speed comparable to the well-known auxiliary function IVA algorithm. The performance of the proposed algorithm is investigated and compared to a benchmark algorithm using real measurements.

Motivation & Objective

• To address the outer permutation ambiguity in Independent Vector Analysis (IVA) for blind source separation in adverse acoustic environments.
• To incorporate spatial prior knowledge (e.g., source direction) into IVA using a Maximum A Posteriori (MAP) framework.
• To develop a fast-converging optimization method that maintains the convergence speed of auxIVA while enforcing spatial constraints.
• To outperform existing spatially constrained IVA methods in terms of interference suppression and computational cost.
• To enable fusion with localization or tracking systems by modeling uncertainty in spatial estimates.

Proposed method

• Derives a MAP formulation of IVA by introducing a Gaussian prior on the demixing filters to encode directional source information.
• Uses the Majorize-Minimize (MM) principle to derive stable, parameter-free update rules that ensure convergence and fast speed.
• Incorporates a geometric constraint (GC) via a prior on the direction of arrival (DOA) to guide the demixing process and resolve permutation ambiguity.
• Applies the STFT to transform time-domain mixtures into frequency bins, enabling frequency-domain IVA with multivariate source priors.
• Implements a filter energy penalty term to stabilize the optimization and prevent numerical issues.
• Uses a free-field model to define the spatial prior, with the prior variance set to σ² = 40 across all frequencies.

Experimental results

Research questions

• RQ1 Can a MAP-based IVA formulation effectively resolve the outer permutation ambiguity using spatial priors without sacrificing convergence speed?
• RQ2 How does the proposed MM-based optimization compare to gradient-based methods in terms of convergence speed and stability?
• RQ3 What is the performance gain of the proposed method in terms of SIR and SDR across varying reverberation conditions?
• RQ4 Can the method maintain high source separation quality while incorporating uncertainty in spatial estimates?
• RQ5 How does the computational cost of the proposed method compare to state-of-the-art spatially constrained IVA algorithms?

Key findings

• GC auxIVA achieves higher Signal-to-Interference Ratio (SIR) than GC gradIVA across all room conditions, indicating superior interference suppression.
• GC auxIVA achieves SDR performance comparable to GC gradIVA and is only slightly lower than auxIVA, despite the added spatial prior.
• The SDR decreases with increasing reverberation time (T60), as expected due to lower Direct-to-Reverberant energy Ratio (DRR).
• GC auxIVA converges significantly faster than GC gradIVA, requiring fewer iterations to reach convergence, despite slightly higher computational cost per iteration.
• The cost function of GC auxIVA and auxIVA converges at nearly identical rates, confirming that the spatial prior does not impair convergence speed.
• The method is computationally more efficient overall than GC gradIVA due to fewer required iterations, even though each iteration is slightly more expensive.