A Unified Bayesian View on Spatially Informed Source Separation and Extraction based on Independent Vector Analysis

Abstract: Signal separation and extraction are important tasks for devices recording audio signals in real environments which, aside from the desired sources, often contain several interfering sources such as background noise or concurrent speakers. Blind Source Separation (BSS) provides a powerful approach to address such problems. However, BSS algorithms typically treat all sources equally and do not resolve uncertainty regarding the ordering of the separated signals at the output of the algorithm, i.e., the outer permutation problem. This paper addresses this problem by incorporating prior knowledge into the adaptation of the demixing filters, e.g., the position of the sources, in a Bayesian framework. We focus here on methods based on Independent Vector Analysis (IVA) as it elegantly and successfully deals with the internal permutation problem. By including a background model, i.e., a model for sources we are not interested to separate, we enable the algorithm to extract the sources of interest in overdetermined and underdetermined scenarios at a low computational complexity. The proposed framework allows to incorporate prior knowledge about the demixing filters in a generic way and unifies several known and newly proposed algorithms using a Bayesian view. For all algorithmic variants, we provide efficient update rules based on the iterative projection principle. The performance of a large variety of representative algorithmic variants, including very recent algorithms, is compared using measured room impulse responses.

• The paper presents a unified Bayesian formulation that integrates spatial and spectral priors into frequency-domain source separation to resolve permutation ambiguities.
• The method leverages MAP estimation with MM optimization, enabling efficient demixing and background modeling in both overdetermined and underdetermined settings.
• The approach demonstrates robust performance with practical improvements in SDR/SIR and reduced iteration counts in real reverberant environments.

Unified Bayesian Framework for Spatially Informed Source Extraction and Separation via Independent Vector Analysis

Introduction and Motivation

This paper presents a rigorous Bayesian formulation for spatially informed source extraction and separation that unifies and generalizes state-of-the-art approaches based on Independent Vector Analysis (IVA). The central innovation is the integration of prior spatial and spectral knowledge into the adaptation of frequency-domain demixing filters within a Maximum A Posteriori (MAP) estimation framework. This approach effectively resolves both the well-known inner and outer permutation ambiguities of frequency-domain Blind Source Separation (BSS) while enabling efficient signal extraction in both overdetermined and underdetermined scenarios, through the explicit modeling of background (BG) sources. The proposed framework subsumes established algorithms such as IVA, ILRMA, and recent spatially constrained and extraction-based variants, and supports the systematic incorporation of informative priors, including free-field directional constraints.

Probabilistic Modeling and Bayesian Formulation

The paper’s probabilistic modeling is comprehensive:

• The observed $M$-microphone $Q$-source convolutive mixture is modeled in the Short-Time Fourier Transform (STFT) domain, allowing for classic frequency-domain separation.
• Demixing is parameterized by frequency-dependent matrices $W_f$, decomposed into two parts: filters for the sources of interest (SOI) and filters for the background (BG) sources.
• The Bayesian estimation framework targets the full posterior $p(W|X)$, integrating evidence from the signal model, background model, and prior distributions on the demixing filters.

A key technical contribution is the explicit derivation of the MAP estimation problem, incorporating:

• Source models: including super-Gaussian, time-varying Gaussian, and Nonnegative Matrix Factorization (NMF)-based models for the SOIs.
• Flexible priors on demixing vectors: quadratic (steering spatial ones or nulls), Euclidean, or generic forms, constructed from free-field models and leveraging array geometry and a priori DOA knowledge.
• Background model: a BG component (for over- or underdetermined scenarios when extracting only a subset of sources) is modeled as independent multivariate complex Gaussian in each TF-bin, which captures broad classes of acoustic interference (e.g., white or diffuse noise) and enables computational savings.

Crucially, the Bayesian approach enables the design of multimodal algorithms: pure BSS (separating all sources, standard IVA/ILRMA), signal extraction (targeted SOI recovery), and hybrid scenarios (overdetermined/underdetermined with BG modeling), all as special cases.

Algorithmic Realization: Majorize-Minimize and Iterative Projection

Optimization of the MAP objective is carried out using the Majorize-Minimize (MM) principle, which ensures monotonic convergence by iteratively optimizing a tractable quadratic upper bound on the cost function:

• The MM upper bound construction follows [Ono, 2011], generalizing it to the informed setting with priors and BG models.
• For unconstrained SOI channels, update rules align with standard Frequency-Domain Iterative Projection (FDIP) as in auxiliary-function-based IVA.
• When priors are present, updates become regularized by spatial or Euclidean constraints, readily enabling directional or null steering in the filter space.
• The framework provides new derivations and analytic solutions for updating BG filter parameters with or without prior constraints—these are shown to be efficient and amenable to fast computation.
• Update rules cover super-Gaussian, TVG, and NMF-based SOI models, with explicit expressions for updating both demixing matrices and (for NMF models) basis/activation coefficients.

A special focus is given to the design and impact of spatial priors: quadratic forms weighted by free-field steering vectors (imposing either spatial “ones” or “nulls”) and their parameterization for robust practical adaptation to real acoustic environments.

Experimental Evaluation and Results

Extensive experiments are conducted with real measured Room Impulse Responses (RIRs) using a $4$-microphone linear array and $8$-source configurations, including challenging over- and underdetermined, reverberant conditions ($T_{60}=0.2\,\mathrm{s}$ and $0.4\,\mathrm{s}$). The evaluation encompasses a rich family of algorithmic variants defined by combinations of:

• Source models (SG, TVG, NMF),
• Use or omission of BG modeling,
• Quadratic versus Euclidean spatial priors,
• Optimization approaches (gradient descent, IP),
• Extraction versus separation operating modes.

Key performance measures (SDR, SIR, SAR improvement, as per [Vincent et al., 2006]), as well as average runtime per iteration, are reported for all variants.

Numerical highlights include:

• NMF-based source models consistently outperform non-NMF baselines in SDR/SIR improvement, especially at low reverberation times.
• Introduction of BG modeling drastically reduces computational cost while maintaining (and sometimes enhancing) separation and extraction performance, even in underdetermined regimes.
• Algorithms with spatial null priors (constrained on all but one channel) are highly sensitive to reverberation and degrade when free-field assumptions break down, while spatial one priors (constraining only the SOI filter) exhibit more robust performance across room conditions.
• The MM/IP-based methods require an order of magnitude fewer iterations (100 vs. > 2000) compared to the gradient-based approach, yielding significant speed-up.
• Parametric studies (number of NMF bases, SOI model shape parameter $\beta$, input SNR) provide empirical guidance on hyperparameter selection; typically, small numbers of bases ($1$–$2$) and $\beta=1$ (Laplacian) are near-optimal.

Theoretical and Practical Implications

The framework offers several advancements:

• It unifies a heterogeneous literature of spatially informed IVA, ILRMA, and extraction-based BSS methods under a single Bayesian, MM-optimized model with clear pathways to incorporate prior geometrical and statistical information.
• Derivations for BG filter updates under arbitrary spatial constraints (including those employing both spatial nulls and ones) generalize and systematize previously heuristic or ad hoc methods.
• The explicit use of BG modeling enables the direct extension of spatially informed IVA to over- and underdetermined cases, a regime challenging for standard IVA/ILRMA.
• The separation of prior design from update rule derivation—supported by explicit optimization—facilitates principled, task-specific engineering of extraction and separation systems.

Contradictory/strong claims: The paper demonstrates that spatial null priors, while theoretically appealing for exclusion of interfering directions, are counterproductive in highly reverberant settings. Furthermore, it asserts that efficient and optimal BG filter updates—critical for extraction in underdetermined conditions—had not previously been systematically derived from MM/IP principles.

Future Directions

This generic Bayesian framework invites several extensions:

• Integration of data-driven or learned priors using measured RIRs or DOA posteriors.
• Real-time adaptation and computational scaling for large microphone arrays or rapidly shifting acoustic contexts.
• Multi-target extraction and semi-supervised or online learning of NMF bases or other components.
• Joint design with dereverberation and noise-robustness mechanisms beyond free-field modeling.

Conclusion

This work synthesizes, extends, and mathematically systematizes spatially informed source separation and extraction in the frequency domain, offering efficient, theoretically well-founded, and practically effective algorithms for real-environment BSS and extraction under a wide variety of mixing scenarios and prior information regimes. It establishes a basis for robust and scalable deployment of IVA-based signal enhancement with transferable spatial and spectral prior knowledge, with demonstrated efficacy on real-world RIR data and a comprehensive suite of performance benchmarks.

Reference: "A Unified Bayesian View on Spatially Informed Source Separation and Extraction based on Independent Vector Analysis" (2001.05958)