Neural acoustic multipole splatting for room impulse response synthesis

TL;DR

This work tackles room impulse response (RIR) synthesis at unseen receiver positions by introducing Neural Acoustic Multipole Splatting (NAMS), a physics-informed neural field that represents sound fields as a sum of learnable neural multipoles with adjustable directivities and emitted signals, conforming to the Helmholtz equation. A pruning strategy densifies multipoles at initialization and progressively removes redundant poles based on the energy of each multipole's signal, yielding compact models with roughly 20–22% of the poles. Empirical results on MeshRIR and Treble Apartment scenes show that NAMS achieves higher accuracy across key acoustic metrics while maintaining fast inference (about 2.1–2.2 ms) and far fewer parameters than strong baselines like AVR. The approach demonstrates that multipole-based representations can more effectively capture complex room acoustics than monopole models, with pruning enabling interpretable and efficient models suitable for real-time spatial audio rendering.

Abstract

Room Impulse Response (RIR) prediction at arbitrary receiver positions is essential for practical applications such as spatial audio rendering. We propose Neural Acoustic Multipole Splatting (NAMS), which synthesizes RIRs at unseen receiver positions by learning the positions of neural acoustic multipoles and predicting their emitted signals and directivities using a neural network. Representing sound fields through a combination of multipoles offers sufficient flexibility to express complex acoustic scenes while adhering to physical constraints such as the Helmholtz equation. We also introduce a pruning strategy that starts from a dense splatting of neural acoustic multipoles and progressively eliminates redundant ones during training. Experiments conducted on both real and synthetic datasets indicate that the proposed method surpasses previous approaches on most metrics while maintaining rapid inference. Ablation studies reveal that multipole splatting with pruning achieves better performance than the monopole model with just 20% of the poles.

Figures (5)

• Figure 1: Overview of NAMS framework. We optimize each neural acoustic multipole position $\mathbf{x}_p$ and train MLPs to predict its emitted signal $s_p(t)$ and spherical harmonic coefficients $b_{nm,p}(t)$. We synthesize the RIR using the receiver position $\mathbf{x}_r$ and the set of neural acoustic multipole positions, emitted signals, and spherical harmonic coefficients.
• Figure 2: Room and source configurations used for the simulations.
• Figure 3: Comparison of spatial magnitude distributions. Overall magnitude distribution averaged over the 1/3 octave band centered at 4 kHz (Scene from MeshRIR).
• Figure 4: Ground-truth vs. generated RIRs. Samples with comparable amplitude errors to Table \ref{['tab:modelcomparison']} are selected from MeshRIR.
• Figure 5: Multipole positions and signal energies with and without pruning. Circles indicate the multipoles inside the room, with darker red indicating higher energy. The black square denotes the source, and the gray dashed square indicates the measurement region.