Abstract
Given a sound field generated by a sparse distribution of impulse image sources, can the continuous 3D positions and amplitudes of these sources be recovered from discrete, bandlimited measurements of the field at a finite set of locations, e.g., a multichannel room impulse response? Borrowing from recent advances in super-resolution imaging, it is shown that this nonlinear, non-convex inverse problem can be efficiently relaxed into a convex linear inverse problem over the space of Radon measures in R3. The linear operator introduced here stems from the fundamental solution of the free-field inhomogenous wave equation combined with the receivers' responses. An adaptation of the Sliding Frank-Wolfe algorithm is proposed to numerically solve the problem off-the-grid, i.e., in continuous 3D space. Simulated experiments show that the approach achieves near-exact recovery of hundreds of image sources using an arbitrarily placed compact 32-channel spherical microphone array in random rectangular rooms. The impact of noise, sampling rate and array diameter on these results is also examined.
Introduction
Mapping the acoustic environment of a room is a fundamental challenge in spatial audio, with applications ranging from the design of advanced hearing aids to the creation of immersive AR/VR experiences. When sound is emitted in an enclosed space, it reflects off surfaces, creating a complex sound field. These reflections can be modeled as "image sources"—virtual sound sources located outside the physical boundaries of the room.
Traditionally, identifying these sources has relied on "on-the-grid" methods, where researchers define a discrete 3D grid and search for sources at fixed points. However, this approach suffers from the curse of dimensionality; achieving 1cm precision in a standard room would require over 100 million grid points. Furthermore, real-world sources rarely align perfectly with grid points (basis mismatch), and standard peak-picking algorithms often fail when reflections arrive in rapid succession. This paper introduces a "gridless" framework that treats source localization as a convex optimization problem over continuous 3D space, bypassing the resolution limits of traditional discrete methods.
Methodology: Super-Resolution in 3D Space
The core innovation of this work is the application of super-resolution mathematical frameworks to acoustic imaging. By reformulating the search for discrete spikes into a convex linear inverse problem over the space of Radon measures, the authors utilize the Beurling-LASSO (BLASSO) framework.
The Linear Operator
The researchers derived a new linear operator, $\Gamma$, based on the fundamental solution of the free-field inhomogeneous wave equation. This operator maps a distribution of sources in continuous space to the bandlimited signals captured by a microphone array. This allows the model to account for the physical behavior of sound waves and the specific response characteristics of the receivers.
Sliding Frank-Wolfe (SFW) Algorithm
To solve the resulting optimization problem, the authors adapted the Sliding Frank-Wolfe algorithm. Unlike standard greedy algorithms that simply add new points to a set, SFW incorporates a "sliding" step. After a new source is identified, the algorithm jointly refines the positions and amplitudes of all currently estimated sources using a local non-convex search (BFGS). This enables the estimated sources to "drift" toward their exact continuous coordinates, achieving sub-grid precision.
Experimental Setup
The method was validated using a large-scale simulation study designed to establish the mathematical limits of the approach:
- Environment: 200 randomly generated "shoe-box" rooms with varying dimensions (lengths/widths 2–10m, heights 2–5m) and wall absorption coefficients.
- Data Generation: Room Impulse Responses (RIRs) and ground-truth image source positions were generated using the Pyroomacoustics library.
- Hardware Simulation: A 32-channel spherical microphone array (modeled after the Eigenmike) was used, with experiments testing various array diameters and sampling rates.
Key Results and Insights
The algorithm was evaluated on the first 50ms of the RIRs, where early reflections are most distinct. The results demonstrate a significant leap in localization performance:
- Unprecedented Precision: In large rooms, the method achieved a mean angular error of just 0.38° and a mean radial error of less than 0.1 mm. This level of accuracy would require a discrete grid four orders of magnitude larger than the one used for initialization.
- High Source Recovery: The algorithm successfully recovered 94.3% of audible image sources in large rooms. The "sliding" refinement step proved crucial, allowing the system to resolve hundreds of sources where previous sparse methods typically failed beyond a few dozen.
- Robustness to Noise: The system maintained high performance at 40 dB PSNR. While performance naturally declined at 30 dB, the recall rate remained high if the success threshold was slightly relaxed (e.g., from 2° to 6° angular error).
- Acoustic Density: A key finding was the impact of "echo density." In smaller rooms, reflections arrive so close together that they overlap significantly. The study found that increasing the sampling rate (to 32 kHz) or the array diameter (to 42 cm) significantly improved recovery in these dense environments.
Conclusion and Future Directions
By moving from discrete grids to continuous Radon measures, this research provides a path toward near-exact 3D acoustic mapping. However, several challenges remain for real-world application. The current model assumes idealized physics, such as frequency-independent reflections and ideal low-pass filters. Real-world materials exhibit complex, frequency-dependent absorption, and non-rectangular room geometries can create occlusions that the current operator does not model.
Future work will focus on extending this gridless framework to the Fourier and spherical-harmonic domains. This evolution will be necessary to handle the complex frequency dependencies of real-world environments and hardware, ultimately bringing high-precision acoustic imaging to consumer devices.
Summary updated