betaFeedback

WorksComputer Science

Graph Neural Network for Wide-Area 6DoF Ambisonic Rendering

Abstract

Wide-area 6DoF rendering from sparse Ambisonic recordings is commonly addressed by interpolating parametric spatial descriptors rather than the waveforms themselves. This preserves the fidelity of a nearby measured signal and avoids phase artifacts caused by mixing distant arrays. However, existing DirAC-based approaches rely on geometric interpolation weights that do not depend on the acoustic scene observed by the microphones. We propose a lightweight Graph Neural Network (GNN) that refines only this interpolation stage. The known Ambisonic arrays and the query position are represented as a graph whose nodes contain local DirAC features and whose edges encode relative distance and directional relations. The network predicts positive query-conditioned weights over the measured arrays, while the target energy, diffuseness, and direction of arrival are reconstructed by an analytical DirAC layer. Experiments on unseen rooms, unseen speech signals, and unseen array configurations show that the proposed weighting improves energy reconstruction and radiation-pattern similarity with respect to baseline interpolation.

Introduction

Immersive audio for virtual and augmented reality requires rendering methods that enable six-degrees-of-freedom (6DoF) where the listener changes both orientation and position. This requirement is also reflected in the MPEG-I Immersive Audio standard for interactive VR/AR audio rendering with 6DoF listener motion [15]. While rotational rendering can be addressed with standard Ambisonic decoding or binaural spatialization, translational motion is more demanding because the sound field must be estimated at positions where no microphone is available. This problem becomes particularly difficult in wide-area scenarios, where the available Ambisonic arrays are sparse and irregularly distributed.

A first class of methods approaches the problem as sound field reconstruction. Wave-expansion, equivalent-source, and kernel-based formulations estimate the pressure field from acoustic measurements by imposing an explicit propagation model [4, 7, 8, 14, 12, 5, 13]. These methods are physically interpretable and can provide accurate reconstructions when their assumptions are satisfied. However, their usable bandwidth and spatial extent are limited by the sampling geometry, by the number of measurements, and by the validity of the underlying acoustic model. These limitations are especially relevant for large listening regions and sparse microphone layouts, which remain central challenges in physics-informed and large-domain sound field estimation [16, 17].

A second direction is to reconstruct a representation that is useful for rendering, rather than the full complex wave field. Parametric spatial audio methods [6, 1, 12, 3] follow this principle by describing the local sound field through perceptually relevant time-frequency (TF) descriptors. Directional Audio Coding (DirAC), for example, represents each TF bin using energy, active intensity, direction of arrival, and diffuseness [2, 3]. These quantities do not preserve the full pressure field, but they capture the directional and direct-to-diffuse structure needed to control a binaural renderer. This makes them attractive when the goal is perceptually plausible navigation, rather than exact numerical reconstruction.

Wide-area 6DoF Ambisonic rendering builds directly on this idea. In the method [1], DirAC parameters are estimated at multiple Ambisonic recording positions, interpolated at the listener position, and then used to adapt the binaural statistics of the closest recording. This avoids directly mixing Ambisonic signals captured at distant positions, which would introduce phase inconsistencies and comb-filtering artifacts. The method is simple, lightweight, and effective enough to have been selected in the MPEG-I 6DoF Immersive Audio Call for Proposals [1, 15]. Its strength is also its main limitation: the interpolation weights are purely geometric and therefore cannot adapt to the acoustic scene observed by the arrays.

This paper keeps the same parametric rendering structure and modifies only the interpolation stage. Instead of replacing DirAC with a black-box predictor, we use a GNN to estimate scene-dependent interpolation weights from the measured DirAC descriptors and the relative geometry of the sensor layout. This is a natural formulation because the input is an unordered set of Ambisonic arrays, the number of available arrays may vary, and the relevant information is mostly relational. Graph networks provide a natural relational inductive bias for this type of data, while message-passing neural networks update node representations by aggregating pairwise interactions and are permutation-equivariant by construction [18, 11].

The proposed model is deliberately constrained. The network does not directly synthesize sound fields and does not regress arbitrary acoustic parameters at the query position. It predicts positive weights over the known arrays. The target energy, diffuseness, and direction of arrival are then obtained through an analytical DirAC reconstruction layer. This keeps the output interpretable and preserves the rendering assumptions of the original wide-area DirAC pipeline, while allowing the interpolation rule to depend on the measured scene.

The contribution of this work is threefold. First, we formulate wide-area DirAC interpolation as a graph-based weighting problem over a sparse Ambisonic array layout. Second, we introduce a coordinate-free feature representation based on relative distances and directional relations, avoiding dependence on a fixed coordinate system or sensor ordering. Third, we show on unseen rooms, unseen speech signals, and unseen array configurations that the learned weights improve energy reconstruction and first-order Ambisonics (FOA) radiation-pattern similarity.

Problem formulation

Consider a distributed Ambisonic sensor network with MM known arrays at positions S={r1,…,rM}⊂R3\mathcal{S}=\{\boldsymbol{r}_1,\ldots,\boldsymbol{r}_M\}\subset\mathbb{R}^3 and a query position q∈R3\boldsymbol{q}\in\mathbb{R}^3 where the sound field must be rendered. The mmth array records a higher-order Ambisonics (HOA) signal xm(f,n)∈CC\boldsymbol{x}_m(f,n)\in\mathbb{C}^{C}, with C=(L+1)2C=(L+1)^2 channels for spherical-harmonic (SH) order LL, where ff and nn denote the frequency-bin and short-time Fourier transform (STFT) frame indices, respectively. Let xmFOA(f,n)∈C4\boldsymbol{x}^{\mathrm{FOA}}_m(f,n)\in\mathbb{C}^{4} denote the FOA subset and let

CmFOA(f,n)=E ⁣[xmFOA(f,n)xmFOA,H(f,n)]\begin{equation} \mathbf{C}^{\mathrm{FOA}}_m(f,n) = \mathbb{E}\!\left[ \boldsymbol{x}^{\mathrm{FOA}}_m(f,n) \boldsymbol{x}^{\mathrm{FOA},H}_m(f,n) \right] \tag*{(1)} \end{equation}

be its smoothed spatial covariance matrix. The standard DirAC descriptors for each (f,n)(f,n) are then computed as

Em=12Tr⁡ ⁣{CmFOA},im=Re⁡ ⁣{[C1,4mC1,2mC1,3m]},ψm=1−∥im∥2Em+ε,um=im∥im∥2+ε,\begin{equation} \begin{aligned} E_m &= \frac{1}{2} \operatorname{Tr}\!\left\{ \mathbf{C}^{\mathrm{FOA}}_m \right\}, \quad \boldsymbol{i}_m = \operatorname{Re}\!\left\{ \begin{bmatrix} C^m_{1,4}\\ C^m_{1,2}\\ C^m_{1,3} \end{bmatrix} \right\}, \\ \psi_m &= 1- \frac{\lVert\boldsymbol{i}_m\rVert_2} {E_m+\varepsilon}, \quad \boldsymbol{u}_m = \frac{\boldsymbol{i}_m} {\lVert\boldsymbol{i}_m\rVert_2+\varepsilon}, \end{aligned} \tag*{(2)} \end{equation}

where we omitted (f,n)(f,n) for readability, Ci,jm=[CmFOA(f,n)]i,jC^m_{i,j} =[\mathbf{C}^{\mathrm{FOA}}_m(f,n)]_{i,j}, Tr⁡(⋅)\operatorname{Tr}(\cdot) denotes the matrix trace, and Re⁡{⋅}\operatorname{Re}\{\cdot\} extracts the real part. Following [1, 2, 3], we define the energy-normalized directional descriptor as

ηm(f,n)=im(f,n)Em(f,n)+ε=(1−ψm(f,n))um(f,n).\begin{equation} \boldsymbol{\eta}_m(f,n) = \frac{\boldsymbol{i}_m(f,n)}{E_m(f,n)+\varepsilon} = \bigl(1-\psi_m(f,n)\bigr)\boldsymbol{u}_m(f,n). \tag*{(3)} \end{equation}

The interpolation of the parameters is implemented in [1] using geometric weights gm(q)g_m(\boldsymbol{q}) from the enclosing simplex and the estimate at the query position is given as

Ebase(f,n;q)=∑m=1Mgm(q)Em(f,n),ηbase(f,n;q)=∑m=1Mgm(q)ηm(f,n).\begin{align} E_{\mathrm{base}}(f,n;\boldsymbol{q}) &= \sum_{m=1}^{M} g_m(\boldsymbol{q}) E_m(f,n), \tag*{(4)} \\ \boldsymbol{\eta}_{\mathrm{base}}(f,n;\boldsymbol{q}) &= \sum_{m=1}^{M} g_m(\boldsymbol{q}) \boldsymbol{\eta}_m(f,n). \tag*{(5)} \end{align}

The corresponding ψbase\psi_{\mathrm{base}} and ubase\boldsymbol{u}_{\mathrm{base}} follow from ηbase\boldsymbol{\eta}_{\mathrm{base}}. A nearby measured recording is then re-rendered so that its binaural statistics match these target parameters [1]. This avoids the comb-filtering artifacts of direct waveform mixing across distant arrays. The remaining problem is to replace gm(q)g_m(\boldsymbol{q}) with scene-dependent weights that still preserve this analytical structure.

Graph model

We build a complete graph G=(V,E)\mathcal{G}=(\mathcal{V},\mathcal{E}) with M+1M+1 nodes. The first MM nodes correspond to the known arrays and the last node corresponds to the query position. A fully connected graph is used so that every node can exchange information with every other node, avoiding a hard neighborhood definition that would depend on a particular layout.

The node state is intentionally compact. For each measured array we use

hm(0)(f,n)=[log⁡Em(f,n),  ψm(f,n),  s(f),  1,  0],\begin{equation} \boldsymbol{h}^{(0)}_m(f,n)=\big[\log E_m(f,n),\; \psi_m(f,n),\; s(f),\; 1,\; 0\big], \tag*{(6)} \end{equation}

where s(f)s(f) is a normalized frequency tag. The graph is evaluated independently for each TF bin (f,n)(f,n). s(f)s(f) makes the predicted weights frequency-conditioned without exchanging information between different frequency bins. This design keeps the model compact, supports parallel frequency-wise inference, and is compatible with the frequency-separable formulation of linear acoustic propagation.

The query node uses hq(0)=[0,0,s(f),0,1]\boldsymbol{h}^{(0)}_q=[0,0,s(f),0,1]. Absolute coordinates are excluded from the node features. This allows the network to operate on different sensor layouts, with different numbers of arrays, without learning a sensor-specific coordinate system or ordering. Geometry and directionality enter through the edges. For nodes ii and jj we define

eij=[dij,  ui⊤r^ij,  uj⊤r^ij,  ui⊤uj],\begin{equation} \boldsymbol{e}_{ij}=\big[d_{ij},\; \boldsymbol{u}_i^\top\hat{\boldsymbol{r}}_{ij},\; \boldsymbol{u}_j^\top\hat{\boldsymbol{r}}_{ij},\; \boldsymbol{u}_i^\top\boldsymbol{u}_j\big], \tag*{(7)} \end{equation}

with dij=∥ri−rj∥2d_{ij}=\|\boldsymbol{r}_i-\boldsymbol{r}_j\|_2 and r^ij=(rj−ri)/(dij+ε)\hat{\boldsymbol{r}}_{ij}=(\boldsymbol{r}_j-\boldsymbol{r}_i)/(d_{ij}+\varepsilon). These terms describe sensor spacing, whether the intensity at node ii points toward node jj, whether the reverse relation is also true, and whether the two local flows are aligned. The representation is coordinate-free, rigid translations and rigid rotations do not affect it. Thus, the model can adapt easily to new configurations. The query node has no measured acoustic direction. We therefore set uq=0\boldsymbol{u}_q=\boldsymbol{0} when constructing edges involving the query.

With Λ\Lambda message-passing layers, the update at layer ℓ∈{1,…,Λ}\ell\in\{1,\ldots,\Lambda\} is

mij(ℓ)=ϕmsg(ℓ) ⁣([hi(ℓ−1),hj(ℓ−1),eij]),ai(ℓ)=∑j≠imij(ℓ),hi(ℓ)=ϕupd(ℓ) ⁣([hi(ℓ−1),ai(ℓ)]).\begin{align} \boldsymbol{m}_{ij}^{(\ell)} &= \phi_{\mathrm{msg}}^{(\ell)} \!\left( [\boldsymbol{h}_i^{(\ell-1)}, \boldsymbol{h}_j^{(\ell-1)}, \boldsymbol{e}_{ij}] \right), \tag*{(8)} \\ \boldsymbol{a}_i^{(\ell)} &= \sum_{j\neq i}\boldsymbol{m}_{ij}^{(\ell)}, \tag*{(9)} \\ \boldsymbol{h}_i^{(\ell)} &= \phi_{\mathrm{upd}}^{(\ell)} \!\left( [\boldsymbol{h}_i^{(\ell-1)}, \boldsymbol{a}_i^{(\ell)}] \right). \tag*{(10)} \end{align}

The final node-wise head is shared across all measured nodes and is conditioned on the final query embedding:

am=ϕout ⁣([hm(Λ),hq(Λ),emq]),wm=softplus⁡(am).\begin{equation} a_m = \phi_{\mathrm{out}} \!\left( [\boldsymbol{h}^{(\Lambda)}_m, \boldsymbol{h}^{(\Lambda)}_q, \boldsymbol{e}_{mq}] \right), \qquad w_m=\operatorname{softplus}(a_m). \tag*{(11)} \end{equation}

The shared head makes the architecture permutation-equivariant and independent of MM. Softplus gives positive but unnormalized weights, which is preferable to a softmax here because the model must be able to change absolute energy. The output weights are converted to DirAC parameters by an analytical reconstruction layer,

E^=∑m=1MwmEm,ψ^=∑m=1Mwmψm∑m=1Mwm,v^=∑m=1MwmEm(1−ψm)um,u^=v^∥v^∥2+ε.\begin{align} \hat{E} &= \sum_{m=1}^{M} w_m E_m, %\label{eq:Ehat} \quad \hat{\psi} = \frac{\sum_{m=1}^{M} w_m \psi_m}{\sum_{m=1}^{M} w_m}, \tag*{(12)} \\ \hat{\boldsymbol{v}} &= \sum_{m=1}^{M} w_m E_m (1-\psi_m)\boldsymbol{u}_m, \qquad \hat{\boldsymbol{u}}=\frac{\hat{\boldsymbol{v}}}{\|\hat{\boldsymbol{v}}\|_2+\varepsilon}. \tag*{(13)} \end{align}

The network therefore learns only how to combine measured nodes. The acoustic descriptors are still generated by closed-form equations.

Training and test procedure

The network is trained offline on simulated Ambisonic room impulse responses from the HARP dataset [9]. HARP provides higher-order Ambisonic RIRs generated with the image-source method over a large set of room geometries, absorption conditions, and source-receiver configurations. In this work only the FOA subset is used, since both the input features and the target DirAC descriptors are computed from first-order components.

Network configuration.

In the reported experiments, the GNN uses Λ=2\Lambda=2 message-passing layers. The functions ϕmsg(ℓ)\phi_{\mathrm{msg}}^{(\ell)} and ϕupd(ℓ)\phi_{\mathrm{upd}}^{(\ell)} are implemented as multilayer perceptrons (MLPs) operating on the concatenated message and update features in the preceding equations. The node representations use a hidden dimension of 4848. The output function ϕout\phi_{\mathrm{out}} is a shared node-wise MLP with scalar output, which produces one log-weight ama_m for each measured node before the softplus transformation. No node-specific parameters are used. The complete model contains 27 12127\,121 trainable parameters. Since the graph is fully-connected, the message-passing cost scales as O(M2)\mathcal{O}(M^2) with the number of measured arrays.

Training

Training signals are obtained by convolving the FOA RIRs with dry speech signals randomly selected from the TIMIT corpus [10]. At each training iteration, the room, the speech excerpts, and the spatial sampling of the receiver grid are randomized. This prevents the model from learning a fixed waveform or a fixed microphone layout, and instead encourages it to learn a scene-dependent interpolation rule for the DirAC parameters.

All signals are sampled at 1616 kHz and transformed to the TF domain using a Hann window of 512512 samples, a hop size of 128128 samples, and an FFT size of 512512. The DirAC parameters defined in § 2 are then extracted from the FOA spatial covariance matrices. Low-energy bins are discarded during training by a simple energy threshold relative to the peak energy of the current example, so that the loss is not dominated by bins where the spatial descriptors are unreliable.

For each training sample, M=5M=5 receiver positions are selected as the known Ambisonic arrays and another receiver position is used as the query. The known nodes provide the measured DirAC descriptors, while the query node provides only its position through the edge features defined in § 3. The network outputs positive, unnormalized weights wmw_m for the known nodes. These weights are then converted into the reconstructed query descriptors by the analytical layer in Eqs. (12)-(13). Therefore, supervision is applied to the reconstructed acoustic quantities, not directly to the weights.

The loss combines energy, direct-energy, and spatial regularization terms:

L=λELE+λρLρ+λregLreg.\begin{equation} \mathcal{L} = \lambda_E \mathcal{L}_E + \lambda_{\rho} \mathcal{L}_{\rho} + \lambda_{\mathrm{reg}} \mathcal{L}_{\mathrm{reg}} . \tag*{(14)} \end{equation}

The energy term is computed in logarithmic scale,

LE=E[(10log⁡10(E^+ε)−10log⁡10(E⋆+ε))2],\begin{equation} \mathcal{L}_E = \mathbb{E}\left[ \left(10\log_{10}(\hat{E}+\varepsilon) -10\log_{10}(E^{\star}+\varepsilon)\right)^2 \right], \tag*{(15)} \end{equation}

where (⋅)⋆(\cdot)^{\star} denotes the ground-truth query descriptor. The direct-energy term compares the direct component proxy ρ=E(1−ψ)\rho=E(1-\psi),

Lρ=E[(E^(1−ψ^)−E⋆(1−ψ⋆))2].\begin{equation} \mathcal{L}_{\rho} = \mathbb{E}\left[ \left(\hat{E}(1-\hat{\psi}) -E^{\star}(1-\psi^{\star})\right)^2 \right]. \tag*{(16)} \end{equation}

Finally, spatial smoothness is encouraged by perturbing the query position with a small random displacement δ\boldsymbol{\delta} and penalizing changes in the predicted weights,

Lreg=E[∥w(q)−w(q+δ)∥22].\begin{equation} \mathcal{L}_{\mathrm{reg}} = \mathbb{E}\left[ \left\| \boldsymbol{w}(\boldsymbol{q}) - \boldsymbol{w}(\boldsymbol{q}+\boldsymbol{\delta}) \right\|_2^2 \right]. \tag*{(17)} \end{equation}

In the reported configuration, λE=0.33\lambda_E=0.33, λρ=0.33\lambda_{\rho}=0.33, and λreg=0.05\lambda_{\mathrm{reg}}=0.05. The query and microphone positions are also jittered by 0.10.1 m during training to improve robustness to small localization errors.

Test

The evaluation is carried out with a Monte-Carlo protocol on unseen rooms and unseen speech material. Specifically, we test the trained model on 2525 rooms that were never used during training. For each room, dry speech excerpts not used in training are selected from TIMIT and convolved with the corresponding FOA RIRs. For each room/audio condition, three different five-array configurations are randomly drawn from the available receiver grid, leading to 7575 independent test configurations. In each configuration, the selected arrays are used as known nodes and all remaining receiver positions are used as queries. The same configurations are used for the baseline [1] and for the GNN, so that the comparison isolates the effect of the interpolation strategy.

Evaluation Metrics

To assess the reconstruction quality at the query q\boldsymbol{q}, we compare the predicted parameters (denoted by ⋅^\hat{\cdot}) against the ground truth (⋅∗\cdot^*) over reliable TF bins. Define

NMSE⁡x=10log⁡10∑f,n(x^(f,n)−x⋆(f,n))2∑f,n(x⋆(f,n))2+ε,x∈{E,ψ}.\begin{equation} \operatorname{NMSE}_x =10\log_{10} \frac{ \sum_{f,n} \left( \hat{x}(f,n)-x^{\star}(f,n) \right)^2 }{ \sum_{f,n} \left( x^{\star}(f,n) \right)^2 +\varepsilon }, \qquad x\in\{E,\psi\}. \tag*{(18)} \end{equation}

where ϵ\epsilon prevents division by zero. Specifically, NMSE⁡E\operatorname{NMSE}_E measures the normalized reconstruction error of the acoustic energy density over the reliable TF bins, whereas NMSE⁡ψ\operatorname{NMSE}_{\psi} measures the error in diffuseness, i.e., in the direct-to-diffuse balance represented by the DirAC model.

Converting parametric DirAC states back into waveform signals is an inherently lossy rendering step that can perceptually smear spatial cues. To assess the actual sound energy density map presented to the listener, we evaluate also the output signal. Let D(θ,ϕ)D(\theta, \phi) be the directional energy map computed by steering the final reconstructed Ambisonic signals toward azimuth θ\theta and elevation ϕ\phi. Let D^i\hat{D}_i and Di⋆D_i^{\star} denote the predicted and ground-truth directional-energy values at grid direction ii. Their weighted structural correlation is

P=∑iΦi(D^i−μ^)(Di⋆−μ⋆)∑iΦi(D^i−μ^)2∑iΦi(Di⋆−μ⋆)2,\begin{equation} P = \frac{ \sum_i \Phi_i (\hat{D}_i-\hat{\mu}) (D_i^{\star}-\mu^{\star}) }{ \sqrt{ \sum_i \Phi_i (\hat{D}_i-\hat{\mu})^2 \sum_i \Phi_i (D_i^{\star}-\mu^{\star})^2 } }, \tag*{(19)} \end{equation}

where μ^\hat{\mu} and μ⋆\mu^{\star} are the corresponding weighted means and Φi∝cos⁡(ϕi)\Phi_i \propto \cos(\phi_i) is a polar weight compensating for the grid distortion at the poles. This metric is bounded between 11 (perfect structural match), 00 (no correlation), and −1-1 (opposite spatial focus).

The reference method of [1], used here as the baseline, constructs a Delaunay triangulation of the array positions. For each query position q\boldsymbol{q}, the corresponding local simplex determines the active arrays, and its barycentric coordinates define the geometric interpolation weights gmbase(q)g_m^{\mathrm{base}}(\boldsymbol{q}). These weights satisfy gmbase(q)≥0,g_m^{\mathrm{base}}(\boldsymbol{q})\geq 0, with ∑m=1Mgmbase(q)=1,\sum_{m=1}^{M}g_m^{\mathrm{base}}(\boldsymbol{q})=1, and zero weight assigned to arrays outside the active simplex. Thus, the baseline estimates are obtained by (4)-(5). Hereafter, the superscript “base\mathrm{base}” identifies quantities obtained with the Delaunay-based baseline.

For comparative analysis, the advantage of the proposed GNN over the Delaunay baseline is defined as

ΔNMSE⁡x=NMSE⁡xbase−NMSE⁡xgnn,ΔP=Pgnn−Pbase.\begin{equation} \Delta\operatorname{NMSE}_x = \operatorname{NMSE}^{\mathrm{base}}_x - \operatorname{NMSE}^{\mathrm{gnn}}_x, \qquad \Delta P = P^{\mathrm{gnn}}-P^{\mathrm{base}}. \tag*{(20)} \end{equation}

Positive values, therefore, indicate an improvement of the GNN.

Results

Figure 1. Spatial distribution of the improvement obtained by the proposed GNN with respect to baseline interpolation, averaged over the Monte-Carlo test set. The panels report ΔNMSEE\Delta\mathrm{NMSE}_E for energy, ΔNMSEψ\Delta\mathrm{NMSE}_{\psi} for diffuseness, and ΔP\Delta P for radiation-pattern similarity. Positive values indicate an improvement of the GNN over baseline interpolation. Black contours mark the zero level.

Figure 1 reports the spatial distribution of the performance difference between the proposed GNN and the baseline. The left panel shows the gain in energy reconstruction. The GNN improves the energy NMSE over a large portion of the evaluated region, with local gains reaching approximately ΔNMSEE≃3\Delta\mathrm{NMSE}_E \simeq 3dB. Negative regions are present but remain spatially and numerically limited, with a maximum improvement of the baseline of ≃0.1\simeq 0.1dB. This indicates that the learned interpolation weights are able to correct part of the mismatch between geometric proximity and actual acoustic variation, which is one of the main limitations of purely Delaunay-based interpolation [1].

The center panel shows the corresponding result for diffuseness. In this case, the GNN does not provide a clear improvement over the baseline, but it closely matches its behavior over most of the domain. The median differences are on the order of 10−310^{-3}, with local variations of about ±5×10−3\pm 5\times10^{-3} and a standard deviation mostly between (3(3 and 5)×10−35)\times10^{-3}, reaching approximately 6×10−36\times10^{-3} only in localized regions. This is consistent with the fact that the baseline is already particularly strong for diffuseness. Unlike energy, diffuseness is bounded in [0,1][0,1] and varies more smoothly after the DirAC analysis, so a geometry-based interpolation of nearby measurements is often sufficient to obtain a reliable estimate. The relevant observation is therefore not that the GNN outperforms the baseline on this descriptor, but that it learns to reproduce a strong geometric estimator without degrading it across unseen rooms, unseen speech signals, and unseen array configurations.

The right panel in Figure 1 reports the variation in radiation-pattern similarity ΔP\Delta P. The GNN achieves a positive gain over most of the domain, with local improvements approaching ΔP≃0.1\Delta P \simeq 0.1. This is a substantial increase: since Pearson similarity is bounded, a gain of 0.10.1 means that the predicted directional pattern is not only rescaled, but is noticeably closer in shape to the reference radiation pattern. The Monte-Carlo variability is also consistent with this interpretation. The standard deviation of ΔP\Delta P is generally between 0.080.08 and 0.120.12, with localized values up to approximately 0.160.16 near boundary regions, where the randomly selected array configurations can make the problem more sensitive to the available spatial support. Thus, the gain is not uniform for every room and array selection, but the median improvement remains positive over almost the entire domain. In other words, the GNN better estimates the angular structure that is later used by the DirAC rendering stage, while retaining a clear advantage across unseen rooms, unseen speech signals, and unseen array configurations.

Figure 2

Figure 2. Example of reconstructed FOA radiation pattern at one query position and frequency band. The ground truth is compared with the baseline and the proposed GNN. Color represents normalized energy in dB.

The qualitative example in Figure 2 confirms the behavior observed in the spatial maps. The baseline produces a smoother and more symmetric radiation pattern, losing part of the asymmetric deformation present in the ground truth. The GNN reconstruction better follows the dominant lobe and the overall directional shape of the reference field. This is consistent with the positive ΔP\Delta P values in Figure 1: the learned weights do not only reduce scalar descriptor errors, but also improve the reconstruction of the directional FOA structure.

These results support the central assumption of the method, replacing fixed geometric weights with scene-dependent graph weights improves the interpolation of DirAC descriptors while preserving the analytical reconstruction structure.

Conclusion

We proposed a graph-based refinement of wide-area DirAC interpolation for 6DoF Ambisonic rendering. The method stays within the parametric rendering pipeline, but replaces the fixed geometric interpolation rule with a scene-dependent graph model. The sensor network is represented explicitly, the feature design is coordinate-free, and the output remains interpretable because the target DirAC parameters are reconstructed analytically from positive node weights. This places the method between classical parametric interpolation and data-driven sound field reconstruction.

The current results show clear gains in energy reconstruction and directional consistency, while diffuseness remains close to the Delaunay baseline. This behavior is meaningful because the baseline is already strong for diffuseness, whereas the learned model improves the quantities for which geometry-only interpolation is more limited. The small model size also makes the approach compatible with fast inference and future scene-specific adaptation as new audio becomes available.

Future work will extend the evaluation beyond simulated RIRs by testing the method on real multi-point Ambisonic measurements. Perceptual experiments, such as MUSHRA-style listening tests, will also be needed to quantify whether the observed improvements in energy and radiation-pattern similarity translate into audible improvements in 6DoF rendering quality. A further extension will investigate cross-frequency modeling. A spectral encoder or autoencoder bottleneck could couple neighboring frequency bins before graph aggregation, allowing the model to exploit spectral continuity while retaining frequency-resolved DirAC reconstruction. Another promising direction is scene-adaptive fine-tuning, where the pre-trained network is further optimized on a short calibration recording from the target scene. This would preserve the generalization provided by the synthetic training set while allowing the graph weights to adapt to the specific room, source distribution, and microphone layout used at test time.

References

  1. [1]Politis, A., Pajunen, L., Leppänen, J., Mate, S., & Eronen, A. (2023). Wide-Area 6DOF Rendering of Multi-Point Ambisonic Recordings Based on Interpolation of Spatial Parameters. Proceedings of the 2023 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). https://doi.org/10.1109/WASPAA58266.2023.10248142
  2. [2]Pulkki, V., Politis, A., Laitinen, M.-V., Vilkamo, J., & Ahonen, J. (2017). First-Order Directional Audio Coding (DirAC). In Parametric Time-Frequency Domain Spatial Audio (pp. 89–140). Wiley. https://doi.org/10.1002/9781119252634.ch5
  3. [3]Politis, A., McCormack, L., & Pulkki, V. (2017). Enhancement of Ambisonic Binaural Reproduction Using Directional Audio Coding with Optimal Adaptive Mixing. IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 379–383. https://doi.org/10.1109/WASPAA.2017.8170059
  4. [4]Pezzoli, M., Miotello, F., Koyama, S., & Antonacci, F. (2025). Low-Rank Adaptation of Deep Prior Neural Networks For Room Impulse Response Reconstruction. Waspaa, 1–4.
  5. [5]McCormack, L., Politis, A., McKenzie, T., Hold, C., & Pulkki, V. (2022). Object-Based Six-Degrees-of-Freedom Rendering of Sound Scenes Captured with Multiple Ambisonic Receivers. Jaes, 70(5), 355–372. https://doi.org/10.17743/JAES.2022.0010
  6. [6]Tylka, J. G., & Choueiri, E. Y. (2020). Fundamentals of a Parametric Method for Virtual Navigation within an Array of Ambisonics Microphones. Journal of the Audio Engineering Society, 68(3), 120–137. https://doi.org/10.17743/jaes.2019.0055
  7. [7]Antonello, N., Sena, E. D., Moonen, M., Naylor, P. A., & Waterschoot, T. van . (2017). Room Impulse Response Interpolation Using a Sparse Spatio-Temporal Representation of the Sound Field. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 25(10), 1929–1941.
  8. [8]Ueno, N., Koyama, S., & Saruwatari, H. (2018). Kernel Ridge Regression with Constraint of Helmholtz Equation for Sound Field Interpolation. Proc. International Workshop on Acoustic Signal Enhancement (IWAENC).
  9. [9]Saini, S., & Peissig, J. (2025). HARP: A Large-Scale Higher-Order Ambisonic Room Impulse Response Dataset. Proceedings of the 2025 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW).
  10. [10]Garofolo, J. S., Lamel, L. F., Fisher, W. M., Fiscus, J. G., Pallett, D. S., & Dahlgren, N. L. (1993). DARPA TIMIT: Acoustic-Phonetic Continuous Speech Corpus CD-ROM. In National Institute of Standards and Technology (Number NISTIR 4930).
  11. [11]Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., & Dahl, G. E. (2017). Neural Message Passing for Quantum Chemistry. Proceedings of the 34th International Conference on Machine Learning, 70, 1263–1272.
  12. [12]Greco, G., Pezzoli, M., Antonacci, F., & Sarti, A. (2026). Parametric Virtual Microphone Techniques for Sound Field Reconstruction with Early Reflection Modeling. EURASIP Journal on Audio, Speech, and Music Processing, 2026(1), 4. https://doi.org/10.1186/s13636-025-00437-y
  13. [13]Miotello, F., Terminiello, F., Pezzoli, M., Bernardini, A., Antonacci, F., & Sarti, A. (2024). A physics-informed neural network-based approach for the spatial upsampling of spherical microphone arrays. Iwaenc, 215–219.
  14. [14]Greco, G., Zhang, D., Pezzoli, M., & Antonacci, F. (2025). Distributed Acoustic Mixture-of-Experts for Spatial Sound Estimation. TechRxiv, 2025(1015). https://doi.org/10.36227/techrxiv.176049751.14301049/v1
  15. [15]ISO/IEC 23090-4:2025 Information technology -- Coded representation of immersive media -- Part 4: MPEG-I immersive audio (Numbers ISO/IEC 23090-4:2025). (2025). ISO/IEC 23090-4:2025 Information technology -- Coded representation of immersive media -- Part 4: MPEG-I immersive audio (Numbers ISO/IEC 23090-4:2025).
  16. [16]Koyama, S., Ribeiro, J. G. C., Nakamura, T., Ueno, N., & Pezzoli, M. (2025). Physics-Informed Machine Learning for Sound Field Estimation: Fundamentals, State of the Art, and Challenges. IEEE Signal Processing Magazine, 41(6), 60–71. https://doi.org/10.1109/MSP.2024.3465896
  17. [17]Figueroa-Duran, A., & Fernandez-Grande, E. (2025). Reconstruction of Reverberant Sound Fields over Large Spatial Domains. Journal of the Acoustical Society of America, 157(1), 180–190. https://doi.org/10.1121/10.0034833
  18. [18]Battaglia, P. W., Hamrick, J. B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski, M., Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., Gulcehre, C., Song, F., Ballard, A., Gilmer, J., Dahl, G. E., Vaswani, A., Allen, K., Nash, C., Langston, V., … Pascanu, R. (2018). Relational Inductive Biases, Deep Learning, and Graph Networks. https://arxiv.org/abs/1806.01261

Paper details

vlnc:W_zef4gtje
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents