Source Localization by Multidimensional Steered Response Power Mapping with Sparse Bayesian Learning

Wei-Ting Lai, Lachlan Birnie, Xingyu Chen, Amy Bastine, Thushara D. Abhayapala, Prasanga N. Samarasinghe

Abstract

We propose an advance Steered Response Power (SRP) method for localizing multiple sources. While conventional SRP performs well in adverse conditions, it remains to struggle in scenarios with closely neighboring sources, resulting in ambiguous SRP maps. We address this issue by applying sparsity optimization in SRP to obtain high-resolution maps. Our approach represents SRP maps as multidimensional matrices to preserve time-frequency information and further improve performance in unfavorable conditions. We use multi-dictionary Sparse Bayesian Learning to localize sources without needing prior knowledge of their quantity. We validate our method through practical experiments with a 16-channel planar microphone array and compare against three other SRP and sparsity-based methods. Our multidimensional SRP approach outperforms conventional SRP and the current state-of-the-art sparse SRP methods for localizing closely spaced sources in a reverberant room.

Introduction

Locating multiple sound sources in noisy, reverberant environments is a significant challenge in acoustic signal processing. While humans can instinctively isolate sounds, algorithms often struggle when sources are positioned close together. The standard approach, Steered Response Power (SRP), is robust but prone to "blurring," where individual sources merge into broad, ambiguous energy peaks. This paper introduces SRP-SBL, a novel method that combines multidimensional SRP mapping with Sparse Bayesian Learning (SBL) to achieve high-resolution source localization.

Methodology

The SRP-SBL framework addresses the limitations of traditional SRP through two primary innovations:

  1. Multidimensional SRP Mapping: Instead of collapsing acoustic data into a single average value, the proposed method represents the SRP map as a multidimensional matrix. This preserves critical information across time frames and frequency bins, providing a richer dataset for optimization.
  2. Multi-dictionary Sparse Bayesian Learning (M-SBL): This statistical framework assumes that sound originates from a few discrete points. Unlike other sparse recovery methods (e.g., OMP or LASSO), M-SBL automatically determines the number of sources by iteratively pruning candidate locations that do not fit the observed data.

To maintain computational efficiency, the system employs a two-grid strategy: a coarse grid for initial mapping followed by a fine grid for high-resolution refinement.

Comparison of SRP maps: (a) conventional SRP-PHAT showing blurred sources and (b) the proposed high-resolution SRP-SBL map.

Experimental Setup

The researchers validated SRP-SBL using speech signals from the Microsoft Scalable Noisy Speech Dataset (MS-SNSD). Real-world testing was conducted in a large meeting room with a reverberation time ($T_{60}$) of approximately 0.7 seconds. A 16-channel planar microphone array (MiniDSP UMA-16 v2) recorded three simultaneous loudspeakers playing human speech.

Results and Discussion

The performance of SRP-SBL was evaluated against conventional SRP-PHAT and other state-of-the-art sparse methods, including M-SBL (applied directly) and SRP-S.

Visual Resolution and Clarity: The visual evidence from the SRP maps highlights the resolution gap. Traditional SRP-PHAT produces a large, ambiguous energy peak that fails to separate sources 1 and 2, which are only 10 degrees apart. In contrast, SRP-SBL generates sharp, localized peaks that align precisely with the true directions of arrival (DoAs). This sharpening effect is a direct result of the sparsity constraint enforced by the SBL framework.

Stability and Accuracy Across Time: A critical metric for real-time systems is the ability to localize sources from short audio segments. As shown in the localization error (LE) analysis, SRP-SBL maintains a consistently low error (below 10 degrees) across all tested durations, from 0.25 to 2.0 seconds. Competing sparse methods like SRP-S show significant instability and high error rates when the audio clip is shorter than one second, making them less reliable for low-latency applications. Furthermore, the multidimensional approach ensures that SRP-SBL remains robust in the presence of the 0.7s reverberation, where standard M-SBL fails to provide accurate estimates.

Localization Error (LE) over time for various methods, demonstrating the stability of SRP-SBL even at short durations.

Conclusion

SRP-SBL offers a significant advancement in acoustic source localization by providing high-resolution mapping in reverberant environments. By preserving time-frequency diversity and leveraging the self-tuning capabilities of Sparse Bayesian Learning, the method achieves superior accuracy without requiring prior knowledge of the source count. Future work will focus on resolving front-back ambiguity in planar arrays and extending the framework to track moving sources.

Summary updated

BibTeX
@inproceedings{lai2024source,
  author = {Wei-Ting Lai and Lachlan Birnie and Xingyu Chen and Amy Bastine and Thushara D. Abhayapala and Prasanga N. Samarasinghe},
  title = {{Source Localization by Multidimensional Steered Response Power Mapping with Sparse Bayesian Learning}},
  booktitle = {IWAENC},
  year = {2024},
  doi = {10.1109/iwaenc61483.2024.10694007},
}
Figure