A Unified SVD-Modal Solution for Sparse Sound Field Reconstruction with Hybrid Spherical-Linear Microphone Arrays

We propose a data-driven sparse recovery framework for hybrid spherical linear microphone arrays using singular value decomposition (SVD) of the transfer operator. The SVD yields orthogonal microphone and field modes, reducing to spherical harmonics (SH) in the SMA-only case, while incorporating LMAs introduces complementary modes beyond SH. Modal analysis reveals consistent divergence from SH across frequency, confirming the improved spatial selectivity. Experiments in reverberant conditions show reduced energy-map mismatch and angular error across frequency, distance, and source count, outperforming SMA-only and direct concatenation. The results demonstrate that SVD-modal processing provides a principled and unified treatment of hybrid arrays for robust sparse sound-field reconstruction.

💡 Research Summary

The paper introduces a unified, data‑driven framework for sparse sound‑field reconstruction using hybrid spherical‑linear microphone arrays (SMA + LMA). The core idea is to treat the combined array as a single acoustic system and to factor its transfer operator (H(f)) by singular value decomposition (SVD): (H = U\Sigma V^{\mathrm H}). (U) contains orthogonal microphone‑domain modes, (V) contains orthogonal field‑domain modes (plane‑wave directions), and (\Sigma) holds the singular values that rank the coupling strength of each mode. By truncating to the (K) largest singular values (the authors test (K = 9, 16, 25), corresponding to SH orders 2–4), the transfer matrix is approximated as (H \approx U_K\Sigma_K V_K^{\mathrm H}). Projecting the measured microphone signals (y) onto (U_K^{\mathrm H}) and whitening by (\Sigma_K^{-1}) produces a well‑conditioned, orthogonal dictionary (\tilde H) and transformed observations (\tilde y). Sparse recovery of the plane‑wave coefficients (x) is then performed with an (\ell_{2,p}) norm ( (0<p<1) ) using an iterative re‑weighted least‑squares (IRLS) algorithm, initialized with an (\ell_1) stage.

When only an SMA is present, the SVD modes coincide with spherical harmonics (SH), showing that the method reduces to conventional SH‑based processing. Adding LMAs enriches the modal basis: the linear arrays contribute highly directional modes that are not captured by SH, especially at low frequencies where spherical Bessel functions are weakly excited. At higher frequencies, SH modes become fully excited and align more closely with the SVD modes, but beyond the spatial‑aliasing limit SH suffers degradation while the SVD basis continues to select stable, data‑driven modes. Principal‑angle analysis between the SH subspace and the SVD subspace confirms a frequency‑dependent divergence, with larger mean angles at low frequencies and convergence at mid‑band.

The experimental evaluation uses a simulated 10 × 8 × 3 m room (RT60 = 0.3 s) generated with MCRoomSim. A hybrid array consists of a 64‑element open SMA (radius 10 cm) surrounded by four 8‑element LMAs placed 0.5 m from the SMA center along the x and y axes. Speech sources (4 s) are placed at distances of 1.5 m, 2.5 m, and 3.5 m, with the number of concurrent sources ranging from 2 to 10. Microphone signals are convolved with room impulse responses and corrupted with spatially white Gaussian noise at 30 dB SNR. The dictionary is built from 642 uniformly sampled directions (icosahedral subdivision). Sparse recovery employs IRLS with dynamic regularization based on diffuseness, starting with 10 iterations of (\ell_1) minimization followed by (\ell_{0.7}) refinement.

Two performance metrics are reported: (1) Energy‑map mismatch (E) (a normalized inner‑product measure) and (2) Angular error (the angular distance between true and estimated source directions). Across all frequencies, the SVD‑modal solutions consistently achieve lower (E) than SMA‑only processing and a naïve joint SMA‑LMA SR, and they are comparable to the previously proposed Residue Refinement (RR) method, which required a two‑stage pipeline. The modal approach also yields comparable or slightly better angular errors, especially when more modes (e.g., (K=25)) are used, indicating that additional weaker modes capture finer spatial detail useful for localization. However, increasing (K) slightly raises the energy‑map mismatch because weaker modes are more susceptible to noise and reverberation, highlighting a trade‑off between map fidelity and localization precision.

The authors conclude that the SVD‑modal framework provides a principled, unified treatment of hybrid SMA–LMA arrays. It generalizes SH processing, introduces well‑conditioned complementary modes, and improves spatial resolution without resorting to ad‑hoc concatenation or heuristic refinement. Future work is suggested on task‑dependent optimal mode selection (e.g., fewer strong modes for energy‑map reconstruction, more modes for high‑precision source localization) and on real‑time implementation and validation in real acoustic environments.