betaFeedback

WorksEngineering

Array-Agnostic High-Order Ambisonics Encoding via Diffusion Posterior Sampling

Abstract

Ambisonics provides a device-independent representation of three-dimensional sound fields; however, extracting these coefficients from practical microphone array recordings remains strongly dependent on array geometry, acoustic transfer functions, and spatial sampling resolution. This pipeline is typically divided into two distinct tasks: encoding, which entails reconstructing the physically resolvable spatial components, and upscaling, which requires generating higher-order coefficients. Model-based approaches often fail on irregular arrays due to ill-conditioned transfer matrices, whereas conventional neural encoders remain strictly tied to the specific acquisition setup seen during training. Furthermore, existing upscaling methods usually assume ideal low-order inputs, ignoring device-specific array artifacts. We present ADEUPS, a unified framework that couples an array-independent Ambisonic diffusion prior with an explicit physical acquisition model during inference, enabling zero-shot encoding and upscaling across diverse array configurations. We train our prior using an order-masking strategy and perform inference via score blending, which we demonstrate to be crucial for solving the underlying inverse problem. Through extensive numerical evaluations across various array geometries, we show that ADEUPS consistently outperforms task- and array-specific baselines. Finally, we conduct subjective listening tests to confirm the perceptual effectiveness of our approach.

Introduction

Spatial audio conveys the directional and reverberant structure of an acoustic scene and is therefore central to immersive communication, media, and extended-reality applications [2]. Among spatial audio representations, Ambisonics is particularly attractive as it represents the sound field through spherical harmonics independently of the device used to capture it [3]. In practice, however, obtaining a desired Ambisonics representation from microphone recordings is constrained by the physical acquisition system. A microphone array observes the sound field through its geometry and acoustic transfer functions, such that limited spatial sampling, microphone placement, diffraction, and steering-matrix ill-conditioning introduce frequency- and array-dependent distortions [4]. Moreover, the finite spatial resolution of the array geometry allows only a subset of the desired spherical-harmonic orders to be reliably extracted. Consequently, practical Ambisonics acquisition entails two coupled reconstruction tasks: (i) encoding, which recovers the spatial components observed by the microphone measurements while mitigating acquisition-induced distortions; and (ii) upscaling, which infers higher-order components that may not be accurately observable by the array. Achieving high-fidelity Ambisonics from practical microphone arrays thus requires addressing both tasks jointly.

Existing approaches have largely addressed these two components of Ambisonics acquisition separately. For encoding, classical signal-independent methods recover the Ambisonics coefficients by inverting a theoretical or measured array steering model, typically regularized to maintain solution stability [5, 6, 7, 8, 9]. While applicable to different array configurations given their physical models, their accuracy deteriorates when the corresponding inverse problem is ill-conditioned or spatially underdetermined. Signal-dependent parametric approaches instead estimate a compact spatial description of the acoustic scene and use it to resynthesize the desired Ambisonics representation [10, 11]. Such methods provide strong physical interpretability, but rely on accurate estimation of spatial parameters and can degrade in reverberant and multi-source environments. Data-driven encoders seek to overcome these limitations by learning the encoding mapping from examples, either by estimating an array-specific linear transform [12, 13] or by directly mapping microphone recordings to Ambisonics coefficients [14]. Recent approaches improve generalization by conditioning the learned mapping on microphone locations [15] or on measured array transfer functions (ATFs) [16]. Nevertheless, the learned mapping remains coupled to the acquisition interface, e.g., through the assumed microphone count or array-dependent training data, limiting its applicability across arbitrary recording configurations.

A complementary line of work addresses Ambisonics upscaling, where higher-order spatial components are inferred from an available lower-order representation. Classical approaches exploit structural models of the sound field, most commonly through sparse plane-wave decomposition and compressed sensing [17, 18, 19, 20, 21], or through parametric spatial analysis and resynthesis [11, 22]. These techniques can effectively recover higher-order information when the assumed spatial structure holds, but become less reliable in complex reverberant scenes where the sound field is not well represented by a small number of directional components. More recently, data-driven approaches have learned the low-to-high-order mapping using multi-stage and hierarchical neural architectures [23, 24, 25], waveform-domain networks [26, 27], and physics-informed learning [28]. Generative models have also recently been explored for recovering the inherently underdetermined higher-order components [29, 30, 31, 32]. However, these approaches generally operate on already-encoded Ambisonics signals and assume that their observed low-order components provide a reliable representation of the underlying sound field. Consequently, they do not account for the array-dependent distortions, spatial aliasing, and estimation errors introduced during the preceding encoding stage. While encoding and upscaling can be applied separately, array-dependent distortions and reconstruction errors introduced during encoding can be carried into the upscaling stage, whose input may differ from the ideal Ambisonics signals used for its design or training. Rather than learning separate mappings for encoding and upscaling, one can formulate the recovery of an ideal high-order Ambisonics (HOA) representation as a single inverse problem directly from microphone measurements. This motivates a joint approach to encoding and upscaling, recovering the sound field despite partial and distorted observations. Our prior work, ADEPS [1], adopted this approach for Ambisonic signals via Diffusion Posterior Sampling (DPS) [33], focusing primarily on refining spatially observable orders, i.e. encoding. Concurrently, Moliner et al. [34] applied a similar formulation to Ambisonics room impulse responses (ARIRs). However, despite the competitive encoding performance of ADEPS, the recovery of higher-order components degrades drastically as the order increases—both objectively and subjectively. Motivated by this limitation, we propose Ambisonics Diffusion Encoding and Upscaling via Posterior Sampling (ADEUPS), a unified, generative, array-agnostic framework for reconstructing HOA representations directly from microphone-array measurements. To this end, we train a diffusion model on ideal HOA using a novel order-masking strategy, which enforces interchannel dependencies and learns a robust sound-field prior independent of the acquisition hardware. At inference, the microphone configuration is captured by an acoustic forward model that samples the target sound field at the physical locations and re-encodes the resulting signals into Ambisonics. Embedded within DPS, this operator, combined with score blending, steers the diffusion reverse process to yield the reconstructed HOA representation. Together, these two contributions allow the same learned prior to compensate for acquisition-induced distortions while inferring spatial components beyond the array’s resolvable order. This separation between the learned prior and the physical observation model enables zero-shot operation across microphone geometries and channel counts without retraining. Extensive numerical evaluations and subjective listening tests demonstrate its effectiveness across diverse array configurations and reconstruction settings, where it achieves state-of-the-art performance—surpassing array-specific and semi-agnostic baselines.

The main contributions are as follows:

  • Array-Agnostic Framework: By utilizing linearly encoded Ambisonics derived from the captured microphone signals, and performing refinement and upscaling entirely within the Ambisonic domain, our generative model decouples the sound-field prior from array geometry, generalizing zero-shot to arbitrary microphone counts and configurations without retraining.

  • Unified Encoding and Upscaling: We solve Ambisonics encoding and order upscaling within a single inverse problem, eliminating error accumulation and performance bottlenecks inherent to cascaded architectures.

  • Physical Guided Posterior Inference: By leveraging the physical measurement operator within test-time posterior sampling, the ideal generative prior offers inductive bias against measurement noise, spatial aliasing, and physical array imperfections by constraining solutions to the manifold of ideal HOA signals.

  • Extensive Experimental Validation: We comprehensively evaluate ADEUPS across diverse across diverse microphone geometries and counts, reconstruction orders, considering both encoding and upscaling settings and comparing against dedicated task- and array-specific baselines. In addition to objective spatial and signal-fidelity measures, we conduct subjective listening tests to assess the perceptual quality of the reconstructed Ambisonics signals.

The rest of this paper is organized as follows. § 2 introduces the Ambisonic signal model, establishes the general problem formulation, and provides the mathematical foundation for DPS. Section III details our proposed framework, where we cast the Ambisonic inverse problem explicitly into the DPS sampling process, describe our order-based masking strategy for intra-order learning, and present our blended score approach alongside a tailored noise scheduler. Section IV presents a comprehensive numerical study evaluating our approach across various array configurations for both encoding and upscaling tasks against task-specific baselines. § 5 reports the results of our subjective listening tests. Finally, Section VI concludes the paper.

Throughout this paper, we denote multi-dimensional quantities using boldface notation, with their dimensions specified upon introduction. Calligraphic letters denote sets, e.g., X\mathcal{X}, with R\mathbb{R} and C\mathbb{C} being the sets of real and complex numbers, while ∥⋅∥\|\cdot\|, ∥⋅∥F\|\cdot\|_{\mathrm{F}}, (⋅)T(\cdot)^{T}, (⋅)H(\cdot)^{H}, and (⋅)†(\cdot)^{\dagger} are the indicator function, ℓ2\ell_{2} norm, Frobenius norm, transpose, conjugate transpose, and Moore-Penrose pseudo-inverse, respectively.

Signal Model and Preliminaries

Signal Model

Ambisonics is a widely used spatial audio representation based on the spherical Fourier transform [3]. In this framework, a continuous sound pressure is decomposed into a series of spherical harmonics (SH) functions of order nn and degree mm, weighted by frequency-dependent Ambisonics coefficients anm(k)a_{nm}(k), where kk denotes the wavenumber. When sampling this sound field using an arbitrary array of QQ omnidirectional microphones, where rq=(rq,θq,ϕq)\mathbf{r}_{q}=(r_{q},\theta_{q},\phi_{q}) specifies the position of the qq-th microphone in spherical coordinates [35], the acoustic pressure captured by the qq-th microphone can be modeled as a linear combination of these coefficients [35].

pq(k)=∑n=0∞∑m=0nbn(k,rq)Ynm(θq,ϕq)anm(k)+νq(k).(1)p_{q}(k)=\sum_{n=0}^{\infty}\sum_{m=0}^{n}b_{n}(k,r_{q})Y_{n}^{m}(\theta_{q},\phi_{q})a_{nm}(k)+\nu_{q}(k). \tag*{(1)}

In (1), bn(k,rq)b_{n}(k,r_{q}) denotes the order-dependent radial function, Ynm(θq,ϕq)Y_{n}^{m}(\theta_{q},\phi_{q}) represents the real-valued SH function, and νq(k)\nu_{q}(k) is zero-mean additive measurement noise, assumed to be uncorrelated with the Ambisonics coefficients and i.i.d. across all microphones.

Because the radial functions bn(k,rq)b_{n}(k,r_{q}) decay rapidly beyond the n>∣krq∣n>\lvert kr_{q}\rvert, the infinite series in (1) is practically truncated at an effective maximum order Neff≥∣k⋅max⁡qrq∣N_{\mathrm{eff}}\geq\lvert k\cdot\max_{q}r_{q}\rvert [35]. This truncation yields a finite set of (Neff+1)2(N_{\mathrm{eff}}+1)^{2} coefficients, enabling a compact matrix representation of the array acquisition process:

p(k)=V(k)aNeff(k)+ν(k)∈CQ,(2)\mathbf{p}(k)=\mathbf{V}(k)\mathbf{a}_{N_{\mathrm{eff}}}(k)+\boldsymbol{\nu}(k)\in\mathbb{C}^{Q}, \tag*{(2)}

where aNeff(k)=[a00(k),…,aNeffNeff(k)]Ta_{N_{\mathrm{eff}}}(k)=[a_{00}(k),\ldots,a_{N_{\mathrm{eff}}N_{\mathrm{eff}}}(k)]^{T} stacks the truncated Ambisonics coefficients, ν(k)∈CQ\boldsymbol{\nu}(k)\in\mathbb{C}^{Q} is the array noise vector, and V(k)∈CQ×(Neff+1)2V(k)\in\mathbb{C}^{Q\times(N_{\mathrm{eff}}+1)^2} is the modal steering matrix. This steering matrix factors into

V(k)=YNeffB(k),(3)V(k)=Y_{N_{\mathrm{eff}}}B(k), \tag*{(3)}

where B(k)∈C(Neff+1)2×(Neff+1)2B(k)\in\mathbb{C}^{(N_{\mathrm{eff}}+1)^2\times(N_{\mathrm{eff}}+1)^2} is a diagonal matrix containing the radial terms and YNeff∈RQ×(Neff+1)2Y_{N_{\mathrm{eff}}}\in\mathbb{R}^{Q\times(N_{\mathrm{eff}}+1)^2} evaluates the spherical harmonics at the microphone coordinates.

Problem Formulation

Ambisonics reconstruction [7] aims to estimate the SH coefficients from discrete microphone array measurements (2). Accordingly, our objective is to reconstruct a desired order-Np≤NeffN_p\leq N_{\mathrm{eff}} Ambisonics representation directly from the microphone-array measurements in (2). We represent such a reconstruction by an operator R\mathrm{R} satisfying

a^Np(k)=R(p(k),V(k)),(4)\hat{a}_{N_p}(k)=\mathrm{R}\left(p(k),V(k)\right), \tag*{(4)}

where a^Np(k)∈C(Np+1)2\hat{a}_{N_p}(k)\in\mathbb{C}^{(N_p+1)^2} denotes the reconstructed Ambisonics coefficients. Unlike conventional encoding, the target order NpN_p need not be restricted to the spatial order that can be reliably recovered through direct inversion of the array model. Consequently, (4) encompasses both encoding, namely the recovery of spatial components supported by the microphone observations, and upscaling, namely the inference of higher-order components that are insufficiently observed by the array.

By (3), the mapping (4) can be formulated as an inverse problem [5, 12]. However, recovering aNp(k)a_{N_p}(k) from (4) gives rise to several challenges:

C1 Noise Amplification: At low frequencies (k→0k\to0), the decay in B(k)B(k) at higher orders nn renders direct matrix inversion severely ill-posed, leading to extreme microphone noise amplification.

C2 Non-Uniform Sensitivity: Irregular or non-ideal spatial arrangements of microphones degrade the condition number of YNeffY_{N_{\mathrm{eff}}}, causing direction- and channel-dependent reconstruction errors.

C3 High-Frequency Spatial Aliasing: At frequencies for which the microphone array does not sufficiently sample the sound field, spatial aliasing introduces frequency- and geometry-dependent distortions into the recoverable Ambisonics coefficients.

C4 Underdetermined High-Order Recovery: When the desired Ambisonics order exceeds the spatial dimensionality supported by the microphone observations, e.g., Q<(Np+1)2Q<(N_p+1)^2 direct reconstruction becomes underdetermined. Conventional encoding must therefore restrict the recovered order, leaving higher-order spatial information unresolved.

C5 Array-Invariance: The encoder should be applicable across array geometries and microphone counts QQ. Although C1C_1–C3C_3 represent physical degradations in Ambisonic encoding while C4C_4 pertains to Ambisonics upscaling, both stem from the same underlying inverse problem. Our objective is thus to establish a unified spatial mapping operator that addresses these challenges across diverse array configurations (C5C_5). To this end, we assume access during training to ideal, array-agnostic NpN_p-order Ambisonic coefficients, formalized as the dataset

D={aNp(i)}i=1∣D∣.(5)\mathcal{D}=\left\{a_{N_p}^{(i)}\right\}_{i=1}^{|\mathcal{D}|}. \tag*{(5)}

Preliminaries

Here we review some basic preliminaries needed to formulate our proposed ADEUPS in the following section. These include conventional linear encoding, as well as some basics in generative diffusion modelling and sampling.

Linear Encoder

1) Linear Encoder: Conventional linear encoding tackles the encoding component of (4) by recovering Ambisonics coefficients directly from the microphone measurements through a frequency-dependent linear operator. For a prescribed order N≤NeffN\leq N_{\mathrm{eff}}, it forms

a^N(k)=EN(k)p(k).(6)\hat{a}_{N}(k)=E_{N}(k)p(k). \tag*{(6)}

Let VN(k)V_N(k) contain the columns of V(k)V(k) in (2) associated with Ambisonics coefficients up to order NN. A common encoding rule is its Tikhonov-regularized pseudoinverse [6, 7]:

EN(k)=VNH(k)[VN(k)VNH(k)+ζ2IQ]−1,(7)E_N(k)=V_N^{H}(k)\left[V_N(k)V_N^{H}(k)+\zeta^2 I_Q\right]^{-1}, \tag*{(7)}

where ζ2\zeta^2 controls the trade-off between inversion accuracy and noise amplification. While this construction applies to any known steering matrix and thus accommodates arbitrary geometries, its output remains constrained by the rank and conditioning of VN(k)V_N(k). Consequently, the maximum achievable order is bounded by Nenc=⌊Q−1⌋N_{\mathrm{enc}}=\lfloor\sqrt{Q}-1\rfloor, which, under C4C_4, cannot upscale to recover the HOA if Nenc<NeffN_{\mathrm{enc}}<N_{\mathrm{eff}}. Increasing ζ\zeta suppresses unstable modes at the expense of model mismatch and spatial bias, while decreasing it amplifies microphone noise. Linear encoding consequently remains also sensitive to challenges C1C_1–C3C_3.

Diffusion Models

2) Diffusion Models: Diffusion models map a complex data distribution pdatap_{\mathrm{data}} to a tractable prior distribution (typically Gaussian) along a continuous trajectory. In the forward process, a clean sample x0∼pdatax_0\sim p_{\mathrm{data}} is progressively corrupted across time t∈[0,Tmax⁡]t\in[0,T_{\max}] according to

xt=x0+σ(t)ϵ,ϵ∼N(0,I),(8)x_t=x_0+\sigma(t)\epsilon,\qquad\epsilon\sim\mathcal{N}(0,I), \tag*{(8)}

where σ(t)\sigma(t) defines the noise schedule. To generate samples, this process is reversed by solving the reverse-time stochastic differential equation (SDE) [36]

dxt=[f(xt,t)−g(t)2∇xtlog⁡pt(xt)]dt+g(t)dwˉ,(9)\mathrm{d}x_t=\left[f(x_t,t)-g(t)^2\nabla_{x_t}\log p_t(x_t)\right]\mathrm{d}t+g(t)\mathrm{d}\bar{w}, \tag*{(9)}

where f(xt,t)f(x_t,t) is referred to as the drift term, g(t)g(t) is the diffusion coefficient, wˉ\bar{w} is a time-reversed Wiener process, and pt(xt)p_t(x_t) represents the marginal density of the corrupted state xtx_t at time tt. The gradient ∇xtlog⁡pt(xt)\nabla_{x_t}\log p_t(x_t) defines the marginal score function.

In practice, a neural network (NN) sprior(xt,σ(t);θ)s_{\mathrm{prior}}(x_t,\sigma(t);\theta) is often trained to estimate this score function. Under the Variance Exploding (VE) formulation of elucidating diffusion models (EDMs) [37], the drift term is removed (f=0f=0) and the diffusion parameter is set to g(t)=2tg(t)=\sqrt{2t}, which directly yields σ(t)=t\sigma(t)=t. EDMs precondition the NN as a denoiser Dθ(xt,σ)D_\theta(x_t,\sigma) according to

Dθ(xt,σ)=cskip(σ)xt+cout(σ)Fθ(cin(σ)xt,cnoise(σ)).(10)D_\theta(x_t,\sigma)=c_{\mathrm{skip}}(\sigma)x_t+c_{\mathrm{out}}(\sigma)F_\theta\left(c_{\mathrm{in}}(\sigma)x_t,c_{\mathrm{noise}}(\sigma)\right). \tag*{(10)}

where FθF_{\theta} is the model to be trained, and the coefficients are given by

cin(σ)=(σ2+σdata2)−1/2,(11)c_{\mathrm{in}}(\sigma) = \left(\sigma^{2} + \sigma_{\mathrm{data}}^{2}\right)^{-1/2}, \tag*{(11)}
cskip(σ)=σdata2σ2+σdata2,(12)c_{\mathrm{skip}}(\sigma) = \frac{\sigma_{\mathrm{data}}^{2}}{\sigma^{2} + \sigma_{\mathrm{data}}^{2}}, \tag*{(12)}
cout(σ)=σ⋅σdataσ2+σdata2,(13)c_{\mathrm{out}}(\sigma) = \frac{\sigma\cdot\sigma_{\mathrm{data}}}{\sqrt{\sigma^{2} + \sigma_{\mathrm{data}}^{2}}}, \tag*{(13)}
cnoise(σ)=14ln⁡(σ).(14)c_{\mathrm{noise}}(\sigma) = \frac{1}{4}\ln(\sigma). \tag*{(14)}

Here, σdata\sigma_{\mathrm{data}} represents the standard deviation of the data distribution, while cnoise(σ)c_{\mathrm{noise}}(\sigma) provides a log-scaled noise embedding. These coefficients normalize the network inputs and output targets across noise levels while preserving a skip connection to the data manifold when σ→0\sigma\to0.

Via Tweedie’s formula [38], the marginal score function is related to the ideal denoiser by

∇xtlog⁡pt(xt)≈sprior(xt,σ(t);θ)=Dθ(xt,σ(t))−xtσ(t)2.(15)\nabla_{x_t} \log p_t(x_t) \approx s_{\mathrm{prior}}(x_t,\sigma(t);\theta) = \frac{D_{\theta}(x_t,\sigma(t))-x_t}{\sigma(t)^2}. \tag*{(15)}

Sample generation can thus be executed by solving the probability flow ordinary differential equation (ODE) [39]

dxt=−σ(t)∇xtlog⁡pt(xt)dσ(t)≈−σ(t)sprior(xt,σ(t);θ)dσ(t),(16)\mathrm{d}x_t = -\sigma(t)\nabla_{x_t}\log p_t(x_t)\mathrm{d}\sigma(t) \approx-\sigma(t)s_{\mathrm{prior}}(x_t,\sigma(t);\theta)\mathrm{d}\sigma(t), \tag*{(16)}

or stochastically by integrating an equivalent SDE solver.

Diffusion Posterior Sampling

Rather than generating samples, this approach can be used to solve inverse problems. Consider an unknown clean sample x0x_0 is observed via

y=A(x0)+ϵy,ϵy∼N(0,σy2I),(17)y = \mathcal{A}(x_0) + \epsilon_y,\qquad\epsilon_y \sim\mathcal{N}(0,\sigma_y^2 I), \tag*{(17)}

where A\mathcal{A} denotes a known, potentially nonlinear degradation operator. Sampling from the posterior distribution p(x0∣y)p(x_0|y) can be achieved by substituting the posterior score function into the probability flow ODE in (16). By applying Bayes’ rule, the posterior score decomposes into

∇xtlog⁡pt(xt∣y)=∇xtlog⁡pt(xt)+∇xtlog⁡pt(y∣xt),(18)\nabla_{x_t}\log p_t(x_t|y)=\nabla_{x_t}\log p_t(x_t)+\nabla_{x_t}\log p_t(y|x_t), \tag*{(18)}

where the first term is supplied by the unconditional diffusion prior and the second term corresponds to the likelihood score which we denote as sLH(y∣xt)=∇xtlog⁡pt(y∣xt)s_{\mathrm{LH}}(y \mid x_t) = \nabla_{x_t}\log p_t(y|x_t).

Because the exact likelihood pt(y∣xt)p_t(y|x_t) is intractable for intermediate states t>0t > 0, it is commonly approximated via the estimated target x^0=Dθ(xt,σ(t))\hat{x}_0 = D_{\theta}(x_t,\sigma(t)) [33], [40]:

∇xtlog⁡pt(y∣xt)≈∇xtlog⁡p(y∣x^0)∝(a)−η(t)∇xt∥y−A(x^0)∥22,(19)\nabla_{x_t}\log p_t(y \mid x_t) \approx\nabla_{x_t}\log p(y \mid\hat{x}_0) \mathrel{\mathop{\propto}^{(a)}} -\eta(t)\nabla_{x_t}\left\lVert y-\mathcal{A}(\hat{x}_0)\right\rVert_2^2, \tag*{(19)}

where (a)(a) follows from the Gaussian observation model in (17) and η(t)\eta(t) is a time-dependent weight controlling the guidance strength. Substituting this approximation together with (15) into the probability flow ODE yields the guided generative process:

dxt=−σ(t)[sprior(xt,σ(t);θ)−η(t)∇xt∥y−A(x^0)∥22]dσ(t).(20)\mathrm{d}x_t = -\sigma(t)\left[s_{\mathrm{prior}}(x_t,\sigma(t);\theta)-\eta(t)\nabla_{x_t}\left\lVert y-\mathcal{A}(\hat{x}_0)\right\rVert_2^2\right]\mathrm{d}\sigma(t). \tag*{(20)}

This gradient steering pushes the sampling trajectory toward reconstructions that remain both realistic under the prior and faithful to the observation yy.

ADEUPS

We next introduce ADEUPS, which formulates the inverse problem in the Ambisonics domain to enable zero-shot encoding and upscaling for arbitrary array configurations using a generative prior. Subsection III-A frames this inverse problem within the DPS framework, Subsection III-C outlines our training scheme with order masking and conditioning, and Subsection III-B details the inference pipeline with scheduled score blending, followed by a discussion in Subsection III-D. All signals are defined in the short-time Fourier transform (STFT) domain; time-frame and frequency-bin indices are omitted for brevity.

DPS Formulation for Ambisonics Reconstruction

Recall that due to C4, the desired Ambisonics order, which is the order of the Ambisonics prior, holds Np≥NencN_p \geq N_{\mathrm{enc}}. Similar to [41], [42], we define our prior in the magnitude-compressed STFT domain. We denote this element-wise compression operator as H(z)=∣z∣αei∠(z)\mathcal{H}(z)=\lvert z\rvert^{\alpha}e^{i\angle(z)}, where α,β>0\alpha,\beta>0 are hyperparameters and ∠(z)\angle(z) denotes the phase angle.

We formulate our inverse problem directly in the Ambisonics domain by defining the observation vector yy as the compressed encoded signal of order NpN_p (matching the prior order). Specifically, writing the observations of (17) as y=H(ENpVNp)+ϵyy=\mathcal{H}(E_{N_p}V_{N_p})+\epsilon_y and substituting the signal acquisition model (2) yields:

y=H(ENpVNpH−1(x0))+ϵy.(21)y=\mathcal{H}\left(E_{N_p}V_{N_p}\mathcal{H}^{-1}(x_0)\right)+\epsilon_y. \tag*{(21)}

Accordingly, we define the forward operator A(⋅)\mathcal{A}(\cdot) in (17), which maps the ideal compressed NpN_p order sample x0x_0 to the corresponding observation domain, as

A(x0)=H(ENpVNpH−1(x0)).(22)\mathcal{A}(x_0)=\mathcal{H}\left(E_{N_p}V_{N_p}\mathcal{H}^{-1}(x_0)\right). \tag*{(22)}

This operator explicitly captures the sequence of linear STFT decompression, spatial sampling by microphones, Ambisonics linear encoding, and STFT compression. By modeling these array-specific degradation artifacts (C1–C4) exclusively within the forward pipeline at inference time, the resulting DPS formulation, once properly trained, supports zero-shot generalization across arbitrary microphone geometries (C5).

Inference

We next describe the inference procedure assuming that the parameters θ\theta of the order-conditioned denoiser Dθ(xt,σ(t),κ)D_{\theta}(x_t,\sigma(t),\kappa) are given; their training is discussed in Subsection III-C. The conditioning variable κ\kappa specifies the highest Ambisonics order provided to the denoiser as trusted context. For an array with QQ microphones, we set this order to Nenc=⌊Q−1⌋N_{\mathrm{enc}}=\lfloor\sqrt{Q}-1\rfloor.

Recall that DPS reconstructs the unknown signal by replacing the unconditional score in the reverse diffusion dynamics with the posterior score in (18). Accordingly, ADEUPS inference follows (20) using

spost(xt,σ(t),y)=sprior(xt,σ(t);θ)+sLH(y∣xt),(23)s_{\mathrm{post}}(x_t,\sigma(t),y)=s_{\mathrm{prior}}(x_t,\sigma(t);\theta)+s_{\mathrm{LH}}(y\mid x_t), \tag*{(23)}

where the prior score promotes consistency with the learned distribution of ideal Ambisonics signals, while the likelihood score enforces consistency with the array observation through the physical forward operator (22). We next specify these two terms.

Order-conditioned prior score: A single full-order diffusion score does not explicitly distinguish between Ambisonics components that can be supported by the array and higher-order components that must be inferred from their lower-order context. We therefore exploit two complementary operating modes of the same order-conditioned denoiser. Evaluating it with κ=Nenc\kappa= N_{\mathrm{enc}} yields the generation score

sgen(xt,σ(t);θ)=Dθ(xt,σ(t),κ=Nenc)−xtσ(t)2,(24)s_{\mathrm{gen}}(x_t,\sigma(t);\theta)=\frac{D_\theta(x_t,\sigma(t),\kappa=N_{\mathrm{enc}})-x_t}{\sigma(t)^2}, \tag*{(24)}

which uses the lower-order Ambisonics components as trusted context for generating the remaining HOA components. Evaluating the denoiser with κ=Np\kappa=N_p gives the refinement score

sref(xt,σ(t);θ)=Dθ(xt,σ(t),κ=Np)−xtσ(t)2,(25)s_{\mathrm{ref}}(x_t,\sigma(t);\theta)=\frac{D_\theta(x_t,\sigma(t),\kappa=N_p)-x_t}{\sigma(t)^2}, \tag*{(25)}

which operates on the complete order-NpN_p representation and promotes full-order denoising and refinement. The mechanism by which these two behaviors are induced during training is detailed in Subsection III-C.

The relative importance of generation and refinement varies along the reverse diffusion trajectory. At high noise levels, the higher-order structure has not yet been established, making generation from the trusted lower-order context particularly important. At low noise levels, the reconstructed HOA structure is already largely formed, and continued generative guidance can disturb its fine details. We therefore employ score blending [43] and define

sprior(xt,σ(t);θ)={λgsgen(xt,σ(t);θ)+λr1sref(xt,σ(t);θ),σ(t)>σth,λr2sref(xt,σ(t);θ),σ(t)≤σth,(26)s_{\mathrm{prior}}(x_t,\sigma(t);\theta)= \begin{cases} \lambda_g s_{\mathrm{gen}}(x_t,\sigma(t);\theta)+\lambda_{r1}s_{\mathrm{ref}}(x_t,\sigma(t);\theta), & \sigma(t)>\sigma_{\mathrm{th}},\\ \lambda_{r2}s_{\mathrm{ref}}(x_t,\sigma(t);\theta), & \sigma(t)\le\sigma_{\mathrm{th}}, \end{cases} \tag*{(26)}

where λg,λr1,λr2≥0\lambda_g,\lambda_{r1},\lambda_{r2}\ge0 control the relative score contributions, and σth\sigma_{\mathrm{th}} is a threshold hyperparameter which is set empirically. Thus, the early stages of sampling jointly establish higher-order structure and refine the current Ambisonics estimate, whereas the final stages (where the noise level is below the threshold) focus on full-order refinement.

Array-dependent likelihood score: Having formed the prior score, we instantiate the likelihood term of the generic DPS construction in (19). Following Tweedie’s relation in (15), the blended score provides the clean-signal proxy

x~0=xt+σ(t)2sprior(xt,σ(t);θ).(27)\widetilde{x}_0=x_t+\sigma(t)^2s_{\mathrm{prior}}(x_t,\sigma(t);\theta). \tag*{(27)}

The corresponding likelihood score is

sLH(y∣xt)=−η(t)∇xt∥y−A(x~0)∥22,(28)s_{\mathrm{LH}}(y\mid x_t)=-\eta(t)\nabla_{x_t}\lVert y-A(\widetilde{x}_0)\rVert_2^2, \tag*{(28)}

where A(⋅)A(\cdot) is defined in (22). Following [40], [44], we normalize the guidance strength according to

η(t)=η′σ(t)∥∇xt∥y−A(x~0)∥22∥2,(29)\eta(t)=\frac{\eta'}{\sigma(t)\lVert\nabla_{x_t}\lVert y-A(\widetilde{x}_0)\rVert_2^2\rVert_2}, \tag*{(29)}

where η′>0\eta'>0 is a constant scaling parameter. Consequently, the specific microphone configuration affects the posterior trajectory only through A(⋅)A(\cdot) in the likelihood score, while the learned prior score remains independent of the acquisition array.

Posterior sampling: Substituting (26) and (28) into the DPS dynamics in (20) yields the ADEUPS reverse trajectory. We discretize it over MM steps using the ρ\rho-spaced noise schedule of [37],

σi=(σmax⁡1/ρ+iM−1(σmin⁡1/ρ−σmax⁡1/ρ))ρ,(30)\sigma_i=\left(\sigma_{\max}^{1/\rho}+\frac{i}{M-1}\left(\sigma_{\min}^{1/\rho}-\sigma_{\max}^{1/\rho}\right)\right)^\rho, \tag*{(30)}

for i=0,…,M−1i=0,\ldots,M-1, with σM=0\sigma_M=0. We initialize the trajectory around the encoded observation as

xTmax⁡∼N(y,σmax⁡2I),(31)x_{T_{\max}}\sim\mathcal{N}(y,\sigma_{\max}^2 I), \tag*{(31)}

where y=H(ENpp)y=\mathcal{H}(E_{N_p}p). We then use the first-order Euler sampler with stochastic churn [37]. Denoting the churn-perturbed state and noise level at iteration ii by x^ti\hat{x}_{t_i} and σ^i\hat{\sigma}_i, respectively, the posterior scores are evaluated at (x^ti,σ^i)(\hat{x}_{t_i},\hat{\sigma}_i) and the state is updated as

xti+1=x^ti−σ^i(σi+1−σ^i)(sprior,i+sLH,i).(32)x_{t_{i+1}}=\hat{x}_{t_i}-\hat{\sigma}_i(\sigma_{i+1}-\hat{\sigma}_i)(s_{\mathrm{prior},i}+s_{\mathrm{LH},i}). \tag*{(32)}

After the final sampling step, applying H−1(⋅)\mathcal{H}^{-1}(\cdot) yields the reconstructed order-NpN_p Ambisonics coefficients. The complete inference procedure, including the stochastic churn operations, is summarized in Algorithm 1 and illustrated in Figure 1.

ADEUPS Pipeline

Figure 1. ADEUPS Pipeline

Algorithm 1: ADEUPS Inference

Input: Signals p\mathbf{p}, Steering matrix V\mathbf{V}, Denoiser Dθ(⋅)D_\theta(\cdot), Hyperparameters {λg,λr1,λr2,η′,σth}\{\lambda_g,\lambda_{r1},\lambda_{r2},\eta',\sigma_{\mathrm{th}}\}, Churn parameters Schurn,Snoise,Smin⁡,Smax⁡S_{\mathrm{churn}}, S_{\mathrm{noise}}, S_{\min}, S_{\max}

  1. Set Nenc=⌊Q−1⌋N_{\mathrm{enc}}=\lfloor\sqrt{Q-1}\rfloor and ENpE_{N_p} via (7)

  2. Compute y=H(E~Npp)\mathbf{y}=\mathcal{H}(\widetilde{E}_{N_p}\mathbf{p})

  3. Sample xTmax⁡∼CN(y,σmax⁡2INp)\mathbf{x}_{T_{\max}}\sim\mathcal{CN}(\mathbf{y},\sigma_{\max}^{2}I_{N_p})

  4. for i=0i=0 to M−1M-1 do /* Stochastic Churn sampling */

  5. γi←min⁡(SchurnM,2−1)⋅I{σi∈[Smin⁡,Smax⁡]}\gamma_i\leftarrow\min\left(\frac{S_{\mathrm{churn}}}{M},\sqrt{2}-1\right)\cdot\mathbb{I}\{\sigma_i\in[S_{\min},S_{\max}]\}

  6. σ^i←(1+γi)σi\hat{\sigma}_i\leftarrow(1+\gamma_i)\sigma_i

  7. Sample ϵi∼CN(0,Snoise2I)\epsilon_i\sim\mathcal{CN}(0,S_{\mathrm{noise}}^{2}I)

  8. x^ti←xti+σ^i2−σi2ϵi\hat{\mathbf{x}}_{t_i}\leftarrow\mathbf{x}_{t_i}+\sqrt{\hat{\sigma}_i^{2}-\sigma_i^{2}}\epsilon_i /* Calculate refinement score */

  9. sref←(Dθ(x^ti,σ^i,κ=Np)−x^ti)/σ^i2\mathbf{s}_{\mathrm{ref}}\leftarrow\left(D_\theta(\hat{\mathbf{x}}_{t_i},\hat{\sigma}_i,\kappa=N_p)-\hat{\mathbf{x}}_{t_i}\right)/\hat{\sigma}_i^{2} /* Blend score if over threshold */

  10. if σi>σth\sigma_i>\sigma_{\mathrm{th}} then

  11. sgen←(Dθ(x^ti,σ^i,κ=Nenc)−x^ti)/σ^i2\mathbf{s}_{\mathrm{gen}}\leftarrow\left(D_\theta(\hat{\mathbf{x}}_{t_i},\hat{\sigma}_i,\kappa=N_{\mathrm{enc}})-\hat{\mathbf{x}}_{t_i}\right)/\hat{\sigma}_i^{2}

  12. sprior←λgsgen+λr1sref\mathbf{s}_{\mathrm{prior}}\leftarrow\lambda_g\mathbf{s}_{\mathrm{gen}}+\lambda_{r1}\mathbf{s}_{\mathrm{ref}}

  13. else

  14. sprior←λr2sref\mathbf{s}_{\mathrm{prior}}\leftarrow\lambda_{r2}\mathbf{s}_{\mathrm{ref}} /* Perform DPS step */

  15. x~0←x^ti+σ^i2sprior\widetilde{\mathbf{x}}_0\leftarrow\hat{\mathbf{x}}_{t_i}+\hat{\sigma}_i^{2}\mathbf{s}_{\mathrm{prior}}

  16. sLH←−η(ti)∇x^ti∥y−A(x~0)∥22\mathbf{s}_{\mathrm{LH}}\leftarrow-\eta(t_i)\nabla_{\hat{\mathbf{x}}_{t_i}}\lVert\mathbf{y}-\mathcal{A}(\widetilde{\mathbf{x}}_0)\rVert_2^2

  17. Update xti+1←x^ti−σ^i(σi+1−σ^i)[sprior+sLH]\mathbf{x}_{t_{i+1}}\leftarrow\hat{\mathbf{x}}_{t_i}-\hat{\sigma}_i(\sigma_{i+1}-\hat{\sigma}_i)[\mathbf{s}_{\mathrm{prior}}+\mathbf{s}_{\mathrm{LH}}]

  18. return H−1(x0)\mathcal{H}^{-1}(\mathbf{x}_0)

Order-Masking Prior Training

The inference procedure in Subsection III-B assumes access to an order-conditioned denoiser Dθ(xt,σ(t),κ)D_\theta(\mathbf{x}_t,\sigma(t),\kappa) that can operate both as a full-order refinement prior and as a conditional generator from a prescribed trusted order κ\kappa. We next describe how these complementary behaviors are induced during training.

Training is carried out exclusively using ideal order-NpN_p Ambisonics signals from the dataset [5]; no microphone signals, steering matrices, or array-dependent encoding operators are used. The resulting learned prior is therefore independent of the acquisition configuration.

To expose the denoiser to different amounts of available spatial information, we randomly mask complete Ambisonics orders during training. Let c∈{0,…,(Np+1)2−1}c\in\{0,\ldots,(N_p+1)^2-1\} denote the Ambisonics channel index and let ℓ(c)=⌊c⌋\ell(c)=\lfloor\sqrt{c}\rfloor denote its corresponding SH order. For each training sample, we draw a trusted order κ∈{1,…,Np}\kappa\in\{1,\ldots,N_p\} according to a categorical distribution with probabilities pκp_\kappa, and define

qℓ=I{ℓ≤κ},ℓ=0,…,Np,(33)q_\ell=\mathbb{I}\{\ell\le\kappa\},\qquad\ell=0,\ldots,N_p, \tag*{(33)}

with I\mathbb{I} the indicator function and with the corresponding channel-wise mask qˉ\bar{q} whose entries are [qˉ]c=qℓ(c)[\bar{q}]_c=q_{\ell(c)}. Thus, all components up to order κ\kappa are treated as trusted, while those of orders ℓ>κ\ell>\kappa are regarded as unavailable.

Given an ideal sample x0\mathbf{x}_0, a noise level σ(t)\sigma(t), and ϵ∼CN(0,I)\epsilon\sim\mathcal{CN}(0,I), we construct the input to the denoiser channel-wise as

x~t=diag⁡{qˉ}x0+σ(t)ϵ.(34)\widetilde{\mathbf{x}}_t=\operatorname{diag}\{\bar{q}\}\mathbf{x}_0+\sigma(t)\epsilon. \tag*{(34)}

Consequently, trusted channels follow the standard diffusion corruption process in (8), whereas channels above order κ\kappa contain only noise. The training target, however, always remains the complete clean order-NpN_p Ambisonics signal x0\mathbf{x}_0. Hence, when κ<Np\kappa<N_p, successful denoising requires the network to infer the masked higher-order components from the available lower-order context, thereby learning the inter-order dependencies required by the generation score in (24). For κ=Np\kappa=N_p, no channels are masked and the model reduces to full-order denoising, as required by the refinement score in (25).

The standard EDM preconditioning in (10) must also be adapted to this masking operation. In particular, directly retaining the skip connection on a masked channel would propagate its pure-noise input to the denoiser output. We therefore define the order-conditioned denoiser as

Dθ(x~t,σ(t),κ)=diag⁡{cskip(σ(t))}x~t+diag⁡{cout(σ(t))}Fθ(cin(σ(t))x~t,cnoise(σ(t)),κ),(35)D_\theta(\widetilde{\mathbf{x}}_t,\sigma(t),\kappa)=\operatorname{diag}\{c_{\mathrm{skip}}(\sigma(t))\}\widetilde{\mathbf{x}}_t+\operatorname{diag}\{c_{\mathrm{out}}(\sigma(t))\}F_\theta(c_{\mathrm{in}}(\sigma(t))\widetilde{\mathbf{x}}_t,c_{\mathrm{noise}}(\sigma(t)),\kappa), \tag*{(35)}

where cin(⋅)c_{\mathrm{in}}(\cdot) and cnoise(⋅)c_{\mathrm{noise}}(\cdot) are given in (11), while

cskip(σ)=σdata2σ2+σdata2qˉ,(36)c_{\mathrm{skip}}(\sigma)=\frac{\sigma_{\mathrm{data}}^2}{\sigma^2+\sigma_{\mathrm{data}}^2}\bar{q}, \tag*{(36)}
cout(σ)=σσdataσ2+σdata2qˉ+σdata(1−qˉ).(37)c_{\mathrm{out}}(\sigma)=\frac{\sigma\sigma_{\mathrm{data}}}{\sqrt{\sigma^2+\sigma_{\mathrm{data}}^2}}\bar{q}+\sigma_{\mathrm{data}}(1-\bar{q}). \tag*{(37)}

For a trusted channel, i.e., [qˉ]c=1[\bar{q}]_c=1, these expressions reduce to the standard EDM coefficients in (11). For a masked channel, [qˉ]c=0[\bar{q}]_c=0, the skip coefficient vanishes, preventing its pure-noise realization from being directly propagated to the output, while [cout]c=σdata[c_{\mathrm{out}}]_c=\sigma_{\mathrm{data}} provides the appropriate scaling for predicting the missing component through FθF_\theta. The trusted order κ\kappa is additionally provided explicitly to FθF_\theta, allowing the same network to distinguish between the different order-availability patterns.

Accordingly, the parameters θ\theta are learned by minimizing the channel-wise preconditioned denoising objective

L(θ)=Ex0,ϵ,τ,κ[1C⋅F⋅T∥Fθ(cin(τ)x~τ,cnoise(τ),κ)−x0−diag⁡{cskip(τ)}x~τcout(τ)∥22],(38)\mathcal{L}(\theta)=\mathbb{E}_{\mathbf{x}_0,\epsilon,\tau,\kappa}\left[\frac{1}{C\cdot F\cdot T}\left\lVert\frac{F_\theta(c_{\mathrm{in}}(\tau)\widetilde{\mathbf{x}}_\tau,c_{\mathrm{noise}}(\tau),\kappa)-\mathbf{x}_0-\operatorname{diag}\{c_{\mathrm{skip}}(\tau)\}\widetilde{\mathbf{x}}_\tau}{c_{\mathrm{out}}(\tau)}\right\rVert_2^2\right], \tag*{(38)}

where C=(Np+1)2C=(N_p+1)^2 is the number of Ambisonics channels, FF and TT denote the numbers of frequency bins and time frames, respectively, τ\tau denotes a sampled diffusion noise level, and vector division is taken element-wise. Through the joint sampling of κ\kappa and τ\tau, a single array-independent denoiser thus learns both behaviors required at inference: for κ=Np\kappa=N_p it models the full order-NpN_p Ambisonics distribution, while for κ<Np\kappa<N_p it learns to complete the higher-order representation from the trusted lower-order components.

Discussion

The proposed ADEUPS construction is tailored to the coupled challenges in Subsection III-B by separating the information supplied by the learned Ambisonics prior from that supplied by the physical acquisition model. In particular, rather than relying solely on inversion of the possibly ill-conditioned steering matrix, the posterior trajectory combines a learned prior over ideal Ambisonics with measurement consistency through A(⋅)\mathcal{A}(\cdot). This provides a data-driven regularization mechanism for the noise amplification and non-uniform sensitivity in C1C_1–C2C_2, while explicitly accounting for the array-dependent acquisition effects underlying spatial aliasing in C3. The order-conditioned generation score exploits learned inter-order dependencies to infer components that are insufficiently determined by the measurements, thereby addressing the high-order recovery challenge in C4, with the likelihood score maintaining consistency with the available physical observations. Crucially, the diffusion prior is trained only on ideal Ambisonics signals, whereas the steering and encoding matrices enter solely through the test-time forward operator. Consequently, a single learned model can be applied to different microphone geometries and channel counts without retraining, addressing C5.

The inference procedure introduces several parameters governing the relative strength and numerical realization of posterior sampling. The coefficients λg\lambda_g, λr1\lambda_{r_1}, and λr2\lambda_{r_2} balance generation and refinement, σth\sigma_{\mathrm{th}} determines when the sampler transitions to refinement-only operation, and η\eta controls the strength of measurement guidance. The noise schedule, number of sampling steps, and stochastic-Churn parameters further determine the numerical accuracy and computational cost of the reverse trajectory, while the distribution pκp_{\kappa} used during training determines the trusted-order configurations to which the denoiser is exposed. These quantities can be selected on held-out validation data and subsequently kept fixed across array configurations. In contrast, the target order NpN_p is specified by the reconstruction problem rather than being an algorithmic hyperparameter. Its relation to the effective acquisition order NeffN_{\mathrm{eff}} nevertheless affects the fidelity of the forward model: when Np<NeffN_p < N_{\mathrm{eff}}, contributions from orders above NpN_p are omitted from (22), resulting in model mismatch due to unmodeled spatial components. The impact of this mismatch is examined empirically in Subsection IV-C3.

The formulation also suggests several directions for further development. First, the current trusted order NencN_{\mathrm{enc}} is determined from the microphone count, whereas in practice observability also depends on frequency, array geometry, and the conditioning of the corresponding steering matrix. Replacing this hard order threshold by frequency- and array-dependent confidence information, potentially through soft order or channel masks, could more accurately distinguish measured from prior-dominated components. Further, the present likelihood assumes that the array steering matrix is accurately known; extending posterior sampling to account for calibration errors or uncertainty in the physical acquisition model would broaden its applicability to practical recording systems. These extensions are left for future study.

Empirical Evaluation

Here, we numerically evaluate ADEUPS1, and compare it with several Ambisonics reconstruction methods. The evaluation is structured as follows: Subsection IV-A outlines the experimental setup and the datasets used for training and evaluation. Subsection IV-B details the methodology, including the model configurations, baselines, and the evaluation metrics, while Subsection IV-C reports the numerical results.

Experimental Setup

Training Data

  1. Training Data: For training, we use order Np=5N_p = 5 ARIRs sampled at 16 kHz, generated using HARP room simulation code [45]. We randomly generate rooms with dimensions [6,10]×[6,10]×[2,3]\left[6,10\right] \times\left[6,10\right] \times\left[2,3\right] m and reverberation times T60∈[0.1,0.5]T_{60} \in[0.1, 0.5] s. The Ambisonics measurement point is randomly positioned within a 1×11 \times1 m region on the xyxy-plane at a height of 1.2 to 1.5 m along the zz-axis. Point sources are positioned with distances to the Ambisonics measurement point randomly sampled from 0.8 to 1.5 m, with direction-of-arrivals (DOAs) also drawn randomly. In total, we generate 2000 rooms with 10 ARIRs per room, yielding 20,000 ARIRs. For the speech dataset, we use VCTK [46], which consists of approximately 44 hours of data. We employ an on-the-fly data loader that randomly crops 2.04 s speech segments per source at runtime and convolves them with the corresponding ARIRs. The number of active speakers is randomly chosen to be 1 or 2.

Evaluation Data

  1. Evaluation Data: To evaluate the models, we generate 100 acoustic scenes following the training framework, with two key differences: (i) the Ambisonics order is set to Neff=15N_{\mathrm{eff}} = 15; and (ii) speech signals are taken from the WSJ0 dataset [47]. We evaluate performance across 13 test array geometries: the 7-microphone Project Aria glasses [48] alongside 4 irregular realizations for each microphone count Q∈{4,5,6}Q \in\{4,5,6\}. For Project Aria, we utilize its measured steering matrix, whereas for the irregular geometries, steering matrices are computed analytically. Following [15], the microphone positions for the irregular arrays are randomly sampled subject to minimum and maximum pairwise distances of 0.02 m and 0.18 m, respectively. Microphone signals are generated by projecting the ground-truth Ambisonics onto each array geometry via (2) and adding white Gaussian noise at an signal-to-noise ratio (SNR) of 50 dB, totaling in 1300 microphone signals.

Methodology

Benchmarks

Methodology

We base our diffusion model on the NCSN++M [49] backbone. We incorporate a fully connected layer for trusted order conditioning, replace additive conditioning with adaptive linear norm [50], and resize the input and output fully connected layers to (Np+1)2(N_p+1)^2 channels, resulting in a model with 31.1 M parameters.

For the STFT, an FFT length of 510 samples, a hop size of 128 samples, and a Hann window were employed, resulting in 256 frequency bins and 256 time frames. The compression parameters were set to α=0.67\alpha= 0.67 and β=3\beta= 3 to achieve a unit standard deviation (σdata=1\sigma_{\mathrm{data}} = 1). During training, the noise schedule is configured with σmax⁡=80\sigma_{\max} = 80, σmin⁡=2×10−3\sigma_{\min} = 2 \times10^{-3}, and ρ=10\rho= 10. The order-masking probability is set to p(κ=Np)=p(κ=2)=0.5p(\kappa= N_p) = p(\kappa= 2) = 0.5. During inference, we adopt the same parameters except for σmax⁡\sigma_{\max}, which is reduced to 4. Sampling is performed with M=90M = 90 steps, resulting in 150 function evaluations, using stochastic churn parameters Schurn=5S_{\mathrm{churn}} = 5, Snoise=1S_{\mathrm{noise}} = 1, Stmin⁡=0S_{t_{\min}} = 0, and Stmax⁡=50S_{t_{\max}} = 50. The score weights are set to λg=λr1=0.7\lambda_g = \lambda_{r_1} = 0.7, λr2=1\lambda_{r_2} = 1, and η′=0.5\eta' = 0.5, with a guidance threshold of σth=0.4\sigma_{\mathrm{th}} = 0.4.

  1. Benchmarks: We compare our framework against both model-based methods and learning-based methods, with the former being array-agnostic by design. To the best of our knowledge, no existing data-driven method supports varying array configurations and sizes, and we thus benchmark against both semi-agnostic and array-specific approaches:

  • Linear Encoder: The encoder ENencE_{N_{\mathrm{enc}}} defined in (7).

  • Parametric Method: Based on [10], we estimate the direct and reverberant components of the sound field and

resynthesize the Ambisonics signal up to the target order. To establish an upper bound, we use oracle DOAs to estimate direction of sound sources.

  • UNet-Encoder: Based on [12], this approach utilizes a U-Net architecture to estimate the encoding matrix ENencE_{\mathrm{Nenc}}. As it is array-specific, we train it on one of the four-microphone arrays from the test set. The network has a size of 7.9M parameters.

  • Gen-A: Based on [15], we employ a U-Net NN that estimates the encoding matrix conditioned on the microphone positions to encode up to the spatially resolvable order. Because the NN is constrained to a fixed channel count, we train it across diverse array geometries with Q=5Q = 5 randomly positioned microphones. This network is of size 10.2M parameters.

  • Diff-Encoder: We train a diffusion model using the NCSN++M architecture to generate an NpN_p-order Ambisonics signal conditioned on the captured microphone signals p(k)\boldsymbol{p}(k). Conditioning is performed by concatenating p(k)\boldsymbol{p}(k) with the corrupted state xt\boldsymbol{x}_t along the channel dimension. This approach is array-specific, and is trained on one of the four-microphone arrays from the test set. The network size in this case is 30.8M parameters.

  • Diff-Upscaler: Inspired by [32], we utilize the NCSN++M architecture to upscale from first-order Ambisonics (FOA) to order NpN_p. We train the model on FOA signals linearly encoded from randomly positioned arrays with Q∈{4,6}Q \in\{4,6\} microphones. This network is of size 30.8M parameters.

A summary of the models used in the evaluation is presented in Table I.

Evaluation Metrics

2) Evaluation Metrics: Let zz and z^\hat{z} denote the reference and reconstructed time-domain Ambisonic signals, respectively. Furthermore, let Zc(t,f)Z_c(t,f) and Z^c(t,f)\hat{Z}_c(t,f) represent their uncompressed STFT coefficients for channel c∈{1,…,C}c \in\{1,\ldots,C\}, with tt and ff denoting the time-frame and frequency-bin indices, respectively. We evaluate performance using the following metrics:

(i) Log-Magnitude Spectral Error: Measures the spectral magnitude fidelity across all channels CC, time frames TT, and frequency bins FF as

LogSE⁡=1CTF∑c=1C∑t=1T∑f=1F∣20⋅log⁡10(∣Zc(t,f)∣∣Z^c(t,f)∣)∣\operatorname{LogSE} = \frac{1}{CTF}\sum_{c=1}^{C}\sum_{t=1}^{T}\sum_{f=1}^{F}\left|20\cdot\log_{10}\left(\frac{|Z_c(t,f)|}{|\hat{Z}_c(t,f)|}\right)\right|

(ii) Mean Spectral Coherence: Evaluates the average normalized cross-spectral coherence between the reference and estimated signals:

MSC⁡=1CF∑c=1C∑f=1F∣∑t=1TZc∗(t,f)Z^c(t,f)∣∑t=1T∣Zc(t,f)∣2∑t=1T∣Z^c(t,f)∣2\operatorname{MSC} = \frac{1}{CF}\sum_{c=1}^{C}\sum_{f=1}^{F}\frac{\left|\sum_{t=1}^{T} Z_c^*(t,f)\hat{Z}_c(t,f)\right|}{\sqrt{\sum_{t=1}^{T}|Z_c(t,f)|^2}\sqrt{\sum_{t=1}^{T}|\hat{Z}_c(t,f)|^2}}

(iii) Signal Invariant Signal Distortion Ratio (SI-SDR): Evaluates overall time-domain fidelity while remaining invariant to global gain mismatches via a scalar projection α∈R\alpha\in\mathbb{R}

SI-SDR⁡=10log⁡10(∥αz∥2∥z^−αz∥2),with α=⟨z^,z⟩∥z∥2(39)\operatorname{SI\text{-}SDR} = 10\log_{10}\left(\frac{\|\alpha z\|^2}{\|\hat{z}-\alpha z\|^2}\right),\qquad\text{with }\alpha=\frac{\langle\hat{z},z\rangle}{\|z\|^2} \tag*{(39)}

Results

We present the results in three parts. First, we compare the proposed method to existing array-agnostic baselines, followed by a comparison against array-specific methods. Finally, we conduct an ablation study to analyze our key design choices.

Full- and Semi-Agnostic Methods Evaluation

1) Full- and Semi-Agnostic Methods Evaluation: Table II presents performance of the array-agnostic methods across different orders and array configurations. For first-order encoding, we compare both model-based approaches with ADEUPS and Gen-A (which is restricted to five-microphone arrays). Our method outperforms all baselines in SI-SDR by a substantial margin and achieves superior spectral error across nearly all configurations, while remaining second-best in coherence.

For orders 2-5, the linear encoder and Gen-A are inapplicable; thus, we also evaluate Diff-Upscaler, which operates on linearly encoded FOA. Because its input lies in the Ambisonics domain rather than the microphone domain, it is array-agnostic in the sense that varying microphone counts do not necessitate retraining. We trained Diff-Upscaler exclusively on four- and six-microphone arrays to evaluate its behavior on two out-of-distribution scenarios. On FOA derived from a five-microphone array, the upscaler performs comparably to the four- and six-microphone cases; this demonstrates that although it was not trained on these specific spatial aliasing artifacts, they closely resemble those of the training arrays, which share an open-sphere topology. Conversely, on the Aria glasses array, the upscaler fails entirely. This highlights a fundamental pitfall of supervised training on specific array configurations: generalization becomes unreliable for geometries that deviate significantly from the training set.

These results highlight the strength of ADEUPS, which achieves superior SI-SDR across all array configurations without having observed any of them during training. This performance gap between diffusion models of identical capacity stems from the underlying distribution being learned: ADEUPS trains on a clean NpN_p-order prior and incorporates array-specific forward operators solely during inference, thereby simplifying optimization. In contrast, the diffusion upscaler must learn a joint distribution entangled with diverse encoding artifacts, substantially increasing training complexity and hindering generalizability.

Array Specific Methods Evaluation

2) Array Specific Methods Evaluation: We next evaluate ADEUPS against array-specific baselines: Diff-Encoder, which directly maps microphone signals to Np=5N_p = 5 order Ambisonics signals, and UNet-Encoder, which estimates an encoding matrix to project the microphone signals to Nenc=1N_{\mathrm{enc}} = 1 order Ambisonics. The evaluation results on a four-microphone2 open array are summarized in Table III.

For first-order Ambisonics, our method outperforms both baselines in SI-SDR, whereas the UNet-Encoder achieves the best performance in terms of coherence and spectral error. For orders 2 through 5, we compare directly with Diff-Encoder; our method achieves higher SI-SDR and comparable coherence, while underperforming the baseline by only a small margin in spectral error. Overall, these results demonstrate the robustness and flexibility of our approach compared to methods trained on a specific geometry. Crucially, when evaluated on an unseen 4-microphone configuration, the baselines fail entirely, whereas our method maintains consistent performance.

MethodParametersInput SignalOutput OrderSupported Arrays
Linear Encoder0MicrophonesNencN_{enc}Arbitrary
Parametric0MicrophonesNpN_{p}Arbitrary
UNet-Encoder7.9MMicrophonesNencN_{enc}Fixed geometry (Q=4Q = 4)
Gen-A10.2MMicrophonesNencN_{enc}Arbitrary geometry (Q=5Q = 5)
Diff-Encoder30.8MMicrophonesNpN_{p}Fixed geometry (Q=4Q = 4)
Diff-Upscaler30.8MFOANpN_{p}Arbitrary
ADEUPS (ours)31.1MMicrophonesNpN_{p}Arbitrary

Table 1. Summary of Models and their Specifications.

OrderMethodSI-SDR [dB] ↑ 4 MicsSI-SDR [dB] ↑ 5 MicsSI-SDR [dB] ↑ 6 MicsSI-SDR [dB] ↑ AriaMean Spec. Error [dB] ↓ 4 MicsMean Spec. Error [dB] ↓ 5 MicsMean Spec. Error [dB] ↓ 6 MicsMean Spec. Error [dB] ↓ AriaCoherence ↑ 4 MicsCoherence ↑ 5 MicsCoherence ↑ 6 MicsCoherence ↑ Aria
1Linear6.547.638.863.409.509.288.3310.90.820.820.830.76
1Param. (oracle)3.733.673.772.285.835.875.856.220.710.700.710.67
1Gen-A-3.75---7.37---0.78--
1ADEUPS10.5211.2713.136.445.615.655.157.420.790.790.830.66
2Param. (oracle)2.412.552.601.506.276.256.236.530.690.680.680.64
2Diff-Upscaler5.716.196.74-3.369.209.038.4020.180.740.740.760.63
2ADEUPS8.349.4911.045.876.336.285.847.860.760.770.810.65
3Param. (oracle)1.802.002.001.006.476.446.416.690.670.660.670.63
3Diff-Upscaler5.305.706.13-6.609.189.028.5020.950.700.700.720.63
3ADEUPS7.588.599.695.296.676.626.238.090.740.750.790.60
4Param. (oracle)1.421.621.640.706.656.616.586.820.660.650.660.62
4Diff-Upscaler4.795.265.64-9.139.219.058.6020.260.680.680.690.54
4ADEUPS6.997.948.824.706.996.956.588.330.730.740.770.62
5Param. (oracle)1.781.391.400.506.776.716.676.910.650.640.650.61
5Diff-Upscaler4.224.825.18-11.239.269.098.6718.820.650.650.670.52
5ADEUPS6.347.238.004.027.247.207.428.520.710.720.760.61

Table 2. Performance Comparison Across Orders and Arrays (bold indicates best performance).

OrderMethodSI-SDR [dB] ↑MSE [dB] ↓Coherence ↑
1UNet-Encoder6.655.730.85
1Diff-Encoder9.605.980.78
1ADEUPS10.835.810.77
2Diff-Encoder7.606.340.74
2ADEUPS8.426.450.76
3Diff-Encoder7.036.570.72
3ADEUPS7.726.750.74
4Diff-Encoder6.526.780.70
4ADEUPS7.177.050.72
5Diff-Encoder6.017.000.68
5ADEUPS6.527.270.71

Table 3. Performance Comparison for a Specific 4 Mic Array2.

Ablation Study

To quantify the individual contributions of our proposed design choices, we conduct an ablation study on the evaluation dataset, reporting the mean metrics across all microphone arrays for each Ambisonics order. Table IV summarizes the ablation sequence where the sampling steps where chosen such that each method had 150 model evaluations.

Method/OrderSI-SDR [dB] ↑ 12345Mean Spec. Error [dB] ↓ 12345Coherence ↑ 12345
ADEPS10.535.793.872.521.305.486.577.478.319.160.760.720.690.660.64
ADEPS + Masking10.387.475.914.643.345.526.246.777.307.930.770.740.710.690.67
ADEPS + Masking + Ref Cond.10.587.295.654.393.205.426.196.797.368.000.790.750.730.700.68
ADEPS + Masking + Gen Cond.0.402.703.884.404.4712.8910.099.128.908.830.590.680.700.710.70
ADEPS + Masking + Score Blending10.044.403.042.621.706.6810.2411.6012.5813.050.750.730.710.700.68
ADEPS + Masking + Score Blending + Scheduler (ADEUPS)11.249.338.377.676.955.616.286.626.957.200.790.770.750.740.72

Table IV. Ablation study evaluating the incremental contributions of order-masking and score-blending. The baseline configuration corresponds to ADEPS [1] (without masking or conditioning, using a first-order stochastic Euler sampler). All configurations are evaluated under an identical budget of 150 model evaluations. Best results are shown in bold.

We define the baseline as our previous framework, ADEPS [1], which performs Ambisonics encoding via DPS using an EDM-based learned prior with a first-order stochastic Euler sampler, omitting both order-masking during training and score-blending during inference. Incorporating HOA masking into training—even without conditioning—substantially improves the SI-SDR for higher-order components while maintaining comparable performance on lower-order ones.

Next, we evaluate models trained with both masking and conditioning, examining the individual effects of conditioning on either the refinement score or the generation score. Conditioning solely on the refinement score yields results similar to the masking-only framework. In contrast, using only the generation score severely degrades performance at lower orders due to the absence of a skip connection. Notably, unlike the other configurations, the generation-only model exhibits a reversed performance trend, where performance improves with higher Ambisonics orders.

This observation motivates a hybrid approach that leverages the strong low-order performance of the refinement score while utilizing the low-order components to guide higher-order generation. However, naive score blending initially yields suboptimal results because the generation score degrades quality at lower noise levels. At these noise levels, high-order components are already established; attempting to generate them primarily from low-order components becomes ineffective and even detrimental. To address this, we empirically define a threshold σth\sigma_{\mathrm{th}}, below which the sampler transitions from score blending exclusively to the refinement score. Figure 2 illustrates the iteration-wise dynamics with and without the scheduler—evaluated on the average SI-SDR for both low-order and high-order components—demonstrating its effectiveness in stabilizing convergence during sampling.

SI-SDR over time for ADEUPS with and without scheduler

Figure 2. SI-SDR over time for ADEUPS with and without scheduler.

We next compare performance across different values of NeffN_{\mathrm{eff}}, evaluating our test set at Neff∈{5,10}N_{\mathrm{eff}} \in\{5, 10\}. Table V reports the results per Ambisonics order, averaged across all microphone arrays. At Neff=5N_{\mathrm{eff}} = 5, we observe peak performance because Np=NeffN_p = N_{\mathrm{eff}}, rendering our forward measurement model (19) exact and thus establishing the upper bound on performance. Beyond this, we observe a noticeable performance drop from Neff=5N_{\mathrm{eff}} = 5 to Neff=10N_{\mathrm{eff}} = 10, but only a minimal change between Neff=10N_{\mathrm{eff}} = 10 and Neff=15N_{\mathrm{eff}} = 15. These findings confirm two key insights: (i) The initial performance degradation stems from the truncated prior order (Np=5N_p = 5). Because the prior cannot represent components beyond NpN_p, it fails to account for energy folding caused by spatial aliasing from higher orders. This behavior is further corroborated by the spectral analysis in Figure 3, which reveals that the performance discrepancy between Neff=5N_{\mathrm{eff}} = 5 and Neff=15N_{\mathrm{eff}} = 15 is concentrated at high frequencies where spatial aliasing is most pronounced; improving performance under such conditions would thus require a higher-order prior. (ii) The marginal difference between Neff=10N_{\mathrm{eff}} = 10 and Neff=15N_{\mathrm{eff}} = 15 demonstrates that energy decays rapidly as the order increases. This indicates that Neff=15N_{\mathrm{eff}} = 15 serves as a valid practical approximation of the infinite sum for speech data, supporting the core assumption of our framework.

Spectral error and coherence for truncation orders $N_{\mathrm{eff}}=5$ and $N_{\mathrm{eff}}=15$

Figure 3. Spectral error and coherence for truncation orders Neff=5N_{\mathrm{eff}}=5 and Neff=15N_{\mathrm{eff}}=15.

Listening Experiment

While Section IV provides objective evaluations, subjective listening experiments remain essential to capture perceptual attributes beyond the scope of numerical metrics. The experimental setup is detailed in § 5-A, followed by the evaluation methodology in § 5-B. Finally, the subjective results are discussed in § 5-C.

Setup

We follow the framework outlined in Section IV-A2 to generate the evaluation data, setting the dry signal duration to 5 seconds. The listening test evaluates three screen configurations across distinct acoustic and array setups:

S1S_1 Single-speaker FOA: Room dimensions 6.8 ×\times 8.2 ×\times 3.0$ m (RT60_{60} = 380 ms, DRR = −0.4-0.4 dB), using a 4-microphone array centered at (2.9, 6.2, 1.7) m. A male speaker is positioned at a distance of 1.3 m with (θ,ϕ)=(90∘,22∘)(\theta,\phi) = (90^\circ, 22^\circ).

S2S_2 Dual-speaker FOA: Same acoustic environment as S1S_1 (RT60_{60} = 380 ms, DRR = −0.4-0.4 dB), using a 5-microphone array centered at (2.9, 6.2, 1.7) m. Two sources are positioned in the room: a male speaker located at 1.3 m with (81∘,45∘)(81^\circ,45^\circ) and a female speaker at 1.2 m with (110∘,−45∘)(110^\circ,-45^\circ).

S3S_3 Single-speaker HOA4: Room dimensions 8.8 ×\times 8.2 ×\times 3.0$ m (RT60_{60} = 450 ms, DRR = 3.2 dB), using a 6-microphone array centered at (3.8, 6.2, 1.7) m. A female speaker is positioned at a distance of 1.3 m with (96∘,−32∘)(96^\circ,-32^\circ).

For each scenario, four stimuli are generated: (i) the ground-truth Ambisonic signal at the target order (FOA for S1S_1–S2S_2, HOA of order 4 for S3S_3); (ii) a linearly encoded FOA signal; (iii) the target-order Ambisonics signal produced by ADEUPS; and (iv) an Ambisonics signal generated by a data-driven baseline. Because existing deep learning baselines cannot simultaneously accommodate arbitrary array geometries and higher target orders, different models are used across scenarios: UNet-Encoder and Gen-A (which take raw microphone signals as input) for S1S_1 and S2S_2, and Diff-Upscaler (which operates on the linearly encoded FOA) for S3S_3.

Methodology

The subjective evaluation was conducted following the Multiple Stimuli with Hidden Reference and Anchor (MUSHRA) methodology [51]. Participants were instructed to rate the overall quality of the stimuli relative to an explicit known reference on a continuous scale from 0 to 100, where 100 denotes no perceptible difference from the reference and lower scores indicate increasing degradation. In each scenario, the ground-truth Ambisonic signal served as the reference (and hidden reference), while the linearly encoded FOA signal served as a low-anchor.

NeffN_{\mathrm{eff}}SI-SDR [dB] ↑Mean Spec. Error [dB] ↓Coherence ↑
123451234512345
514.6811.279.929.128.375.055.776.146.476.730.850.830.810.800.78
1011.299.388.417.737.015.606.276.626.957.200.800.780.760.740.72
1511.249.338.377.676.955.616.286.626.957.200.790.770.750.740.72

Table 5. Effect of NeffN_{\mathrm{eff}} on Performance per Order.

A total of 11 listeners participated in the experiment. Prior to testing, each subject completed a brief training session to familiarize themselves with the range of sound quality and the test interface. The presentation order of the test screens and stimuli within each screen was randomized to mitigate potential bias.

Results

The MUSHRA test results are illustrated in Figure 4. For each test screen, a two-way repeated-measures analysis of variance (ANOVA) was conducted to evaluate the impact of the processing method on subjective quality ratings.

MUSHRA results across test screens

Figure 4. MUSHRA results across test screens: Screen 1 (4-microphone array), Screen 2 (5-microphone array), and Screen 3 (6-microphone array). Horizontal lines denote medians, and diamonds denote means.

Screen 1 (S1S_1): The analysis revealed a statistically significant main effect for the processing method (F(3,30)=22.31F(3, 30) = 22.31, p<.001p < .001, ηp2=0.64\eta_p^2 = 0.64), indicating that method choice substantially influenced perceived quality. Post-hoc pairwise comparisons with Bonferroni corrections were performed across the ADEUPS, ground-truth FOA, and UNet-Encoder conditions, which achieved mean scores of 100, 88.72, and 70.90, respectively. No statistically significant difference was observed between ADEUPS and either the ground-truth signal (p=0.10p = 0.10, g=1.16g = 1.16) or the UNet-Encoder (p=0.08p = 0.08, g=1.20g = 1.20). However, ratings for the ground-truth signal were significantly higher than those for the UNet-Encoder (p<0.05p < 0.05, g=2.60g = 2.60).

Screen 2 (S2S_2): Mauchly’s test indicated a violation of sphericity (χ(5)2=15.35\chi^2_{(5)} = 15.35, p=.009p = .009), requiring Greenhouse-Geisser corrections to the degrees of freedom. The main effect of the processing method remained statistically significant (F(1.88,18.78)=40.26F(1.88, 18.78) = 40.26, p<.001p < .001, ηp2=0.72\eta_p^2 = 0.72). Mean scores for ADEUPS, ground-truth FOA, and the UNet-Encoder were 100, 96.09, and 62.68, respectively. Post-hoc Bonferroni-corrected comparisons revealed no significant difference between ADEUPS and the ground-truth signal (p=0.58p = 0.58, g=0.75g = 0.75), whereas Gen-A performed significantly worse than both ADEUPS (p<0.05p < 0.05, g=−2.17g = -2.17) and the ground truth (p<0.05p < 0.05, g=−2.58g = -2.58).

Screen 3 (S3S_3): Similarly, a highly significant main effect of the processing method was observed (F(3,30)=84.94F(3, 30) = 84.94, p<.001p < .001, ηp2=0.84\eta_p^2 = 0.84). The mean scores for ADEUPS, ground-truth FOA, and the Diff-Upscaler were 100, 91.45, and 81.45, respectively. Post-hoc comparisons with Bonferroni correction showed no statistically significant difference between ADEUPS and either the ground-truth signal or the Diff-Upscaler; however, the ground-truth signal rated significantly higher than the Diff-Upscaler.

Conclusions

In this work, we introduced ADEUPS, a novel, array-agnostic framework for Ambisonic encoding and spatial upscaling. By formulating the reconstruction task as a diffusion-based inverse problem solved via DPS, our model learns an acquisition-independent sound-field prior of ideal HOA signals. At inference, array-specific configurations are incorporated into the forward operator to invert device-induced physical artifacts. To further enhance reconstruction quality, we introduced two key innovations: order-masking during training to enforce interchannel dependencies, and score-blending during inference to guide sampling trajectory. Objective and subjective evaluations demonstrate that ADEUPS achieves performance comparable to—and in many cases exceeding—both array-agnostic model-based baselines and data-driven methods trained on specific array topologies. Crucially, ADEUPS operates in a zero-shot manner, requiring no exposure to specific array geometries during training. Future work will explore real-time algorithmic optimizations and extend the framework to dynamic acoustic scenes.

References

  1. [1]A. Milstein, N. Shlezinger, and B. Rafaely, “Array-agnostic ambisonics encoding via diffusion posterior sampling,” arXiv preprint arXiv:2608.24558, 2026.arxiv.org/abs/2608.24558
  2. [2]H. Hachabiboglu et al., “Perceptual spatial audio recording, simulation, and rendering: An overview of spatial-audio techniques based on psychoacoustics,” vol. 34, no. 3, pp. 36–54, 2017.
  3. [3]F. Zotter and M. Frank, Ambisonics: A practical 3D audio theory for recording, studio production, sound reinforcement, and virtual reality. Springer Nature, 2019.DOI
  4. [4]B. Rafaely, “Analysis and design of spherical microphone arrays,” vol. 13, no. 1, pp. 135–143, 2005.DOI
  5. [5]A. Politis and H. Gamper, “Comparing modeled and measurement-based spherical harmonic encoding filters for spherical microphone arrays,” in IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2017, pp. 224–228.DOI
  6. [6]S. Moreau, J. Daniel, and S. Bertet, “3d sound field recording with higher order ambisonics—objective measurements and validation of a 4th order spherical microphone,” in 120th Convention of the AES, 2006, pp. 20–23.
  7. [7]M. A. Poletti, “Three-dimensional surround sound systems based on spherical harmonics,” Journal of the Audio Engineering Society, vol. 53, no. 11, pp. 1004–1025, 2005.
  8. [8]A. Bastine, L. Birnie, T. D. Abhayapala, P. Samarasinghe, and V. Tourbain, “Ambisonics capture using microphone arrays on head-worn device of arbitrary geometry,” in 2022 30th European Signal Processing Conference (EUSIPCO). IEEE, 2022, pp. 309–313.DOI
  9. [9]Y. Gayer, V. Tourbain, Z. Ben-Hur, D. Alon, and B. Rafaely, “Ambisonics encoder for wearable array with improved binaural reproduction,” arXiv preprint arXiv:2507.04108, 2025.arxiv.org/abs/2507.04108
  10. [10]L. McCormack, A. Politis, R. Gonzalez, T. Lokki, and V. Pulkki, “Parametric ambisonic encoding of arbitrary microphone arrays,” vol. 30, pp. 2062–2075, 2022.
  11. [11]A. Politis, S. Tervo, and V. Pulkki, “COMPASS: Coding and multidirectional parameterization of ambisonic sound scenes,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 6082–6086.
  12. [12]M. Heikkinen, A. Politis, and T. Virtanen, “Neural ambisonics encoding for compact irregular microphone arrays,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 701–705.arxiv.org/abs/2401.05916
  13. [13]T. Deppisch, Y. Gao, M. Mittal, B. Stahl, C. Hold, D. Alon, and Z. Ben-Hur, “Residual learning for neural ambisonics encoders,” arXiv preprint arXiv:2601.18322, 2026.arxiv.org/abs/2601.18322
  14. [14]Y. Qiao, V. Kothapally, M. Yu, and D. Yu, “Neural ambisonic encoding for multi-speaker scenarios using a circular microphone array,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025.arxiv.org/abs/2409.06954
  15. [15]M. Heikkinen, A. Politis, K. Drossos, and T. Virtanen, “Gen-A: Generalizing ambisonics neural encoding to unseen microphone arrays,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025.arxiv.org/abs/2501.08047
  16. [16]“Beyond omnidirectional: Neural ambisonics encoding for arbitrary microphone directivity patterns using cross-attention,” in ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 22 587–22 591.
  17. [17]A. Wabnitz, N. Epain, A. McEwan, and C. Jin, “Upscaling ambisonic sound scenes using compressed sensing techniques,” in IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2011.
  18. [18]A. Wabnitz, N. Epain, and C. T. Jin, “A frequency-domain algorithm to upscale ambisonic sound scenes,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2012, pp. 385–388.DOI
  19. [19]N. Epain and C. T. Jin, “Super-resolution sound field imaging with sub-space pre-processing,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2013, pp. 350–354.DOI
  20. [20]G. Routray and R. M. Hegde, “Sparse plane-wave decomposition for upscaling ambisonic signals,” in IEEE International Conference on Signal Processing and Communications (SPCOM), 2020.DOI
  21. [21]M. Kentgens, S. Al Hares, and P. Jax, “On the upscaling of higher-order ambisonics signals for sound field translation,” in European Signal Processing Conference (EUSIPCO). IEEE, 2021, pp. 81–85.DOI
  22. [22]A. Politis, J. Vilkamo, and V. Pulkki, “Sector-based parametric sound field reproduction in the spherical harmonic domain,” vol. 9, no. 5, pp. 852–866, 2015.DOI
  23. [23]G. Routray, S. Basu, P. Baldev, and R. M. Hegde, “Deep-sound field analysis for upscaling ambisonic signals,” in EAA Spatial Audio Signal Processing Symposium, 2019.DOI
  24. [24]S. Gao, J. Lin, X. Wu, and T. Qu, “Sparse DNN model for frequency expanding of higher order ambisonics encoding process,” vol. 30, pp. 1124–1135, 2022.DOI
  25. [25]E. Chatzimoustafa and P. Jax, “Higher-order ambisonics upscaling using gated recurrent units,” in European Signal Processing Conference (EUSIPCO). IEEE, 2025, pp. 171–175.
  26. [26]Y. Luo and N. Mesgarani, “Conv-tasnet: Surpassing ideal time–frequency magnitude masking for speech separation,” vol. 27, no. 8, pp. 1256–1266, 2019.DOI
  27. [27]I. Nawfal et al., “Ambisonics super-resolution using a waveform-domain neural network,” in AES Conference on Audio for Virtual and Augmented Reality, 2024.arxiv.org/abs/2508.00240
  28. [28]F. Miotello, F. Terminiello, M. Pezzoli, A. Bernardini, F. Antonacci, and A. Sarti, “A physics-informed neural network-based approach for the spatial upsampling of spherical microphone arrays,” in IEEE International Workshop on Acoustic Signal Enhancement (IWAENC), 2024, pp. 215–219.arxiv.org/abs/2407.18732
  29. [29]L. Zhang, X. Wang, R. Hu, D. Li, and W. Tu, “Optimization of sound fields reproduction based higher-order ambisonics (HOA) using the generative adversarial network (GAN),” Multimedia Tools and Applications, vol. 80, no. 2, pp. 2205–2220, 2021.DOI
  30. [30]A. Milstein, N. Shlezinger, and B. Rafaely, “DiffAU: Diffusion-based ambisonics upscaling,” 2026.arxiv.org/abs/2510.00180
  31. [31]E. Picard, D. Di Carlo, A. A. Nugraha, M. Fontaine, and K. Yoshii, “SIRUP: A diffusion-based virtual unmixer of steering vectors for highly-directive spatialization with first-order ambisonics,” arXiv preprint arXiv:2602.17732, 2026.arxiv.org/abs/2602.17732
  32. [32]A. Milstein, N. Shlezinger, and B. Rafaely, “Generative learning for ambisonic upscaling,” 2026. [Online]. Available: https://arxiv.org/abs/2609.23479
  33. [33]H. Chung, J. Kim, M. T. McCann, M. L. Klasky, and J. C. Ye, “Diffusion posterior sampling for general noisy inverse problems,” in International Conference on Learning Representations (ICLR), 2022.arxiv.org/abs/2209.14687
  34. [34]E. Moliner, C. Hold, J. A. Ortiz, S. Prepelita, N. Ananthabhotla, D. Wong, S. Parekh, and S. Lee, “Ambisonics encoding of room impulse responses using a device-agnostic diffusion model,” arXiv preprint arXiv:2608.14097, 2026.arxiv.org/abs/2608.14097
  35. [35]B. Rafaely, Fundamentals of spherical array processing. Springer, 2015, vol. 8.
  36. [36]Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” arXiv preprint arXiv:2011.13456, 2020.arxiv.org/abs/2011.13456
  37. [37]T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,” Advances in neural information processing systems, vol. 35, pp. 26 565–26 577, 2022.arxiv.org/abs/2206.00364
  38. [38]H. E. Robbins, “An empirical bayes approach to statistics,” in Breakthroughs in Statistics: Foundations and basic theory. Springer, 1992, pp. 388–394.
  39. [39]Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” arXiv preprint arXiv:2011.13456, 2020.arxiv.org/abs/2011.13456
  40. [40]E. Moliner, J. Lehtinen, and V. Välimäki, “Solving audio inverse problems with a diffusion model,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023.arxiv.org/abs/2210.15228
  41. [41]T. Gerkmann and R. Martin, “Empirical distributions of dft-domain speech coefficients based on estimated speech variances,” in Proc. Int. Workshop Acoust. Echo Noise Control, 2010.
  42. [42]J. Richter, S. Welker, J.-M. Lemercier, B. Lay, and T. Gerkmann, “Speech enhancement and dereverberation with diffusion-based generative models,” vol. 31, pp. 2351–2364, 2023.DOI
  43. [43]N. Liu, S. Li, Y. Du, A. Torralba, and J. B. Tenenbaum, “Compositional visual generation with composable diffusion models,” in European conference on computer vision. Springer, 2022, pp. 423–439.arxiv.org/abs/2206.01714
  44. [44]S. Kim, H. Kim, and S. Yoon, “Guided-tts 2: A diffusion model for high-quality adaptive text-to-speech with untranscribed data,” arXiv preprint arXiv:2205.15370, 2022.arxiv.org/abs/2205.15370
  45. [45]S. Saini and J. Peissig, “HARP: A large-scale higher-order ambisonic room impulse response dataset,” in IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), 2025.arxiv.org/abs/2411.14207
  46. [46]J. Yamagishi, C. Veaux, and K. MacDonald, “CSTR VCTK Corpus,” 2019.
  47. [47]J. S. Garofolo, D. Graff, D. B. Paul, and D. S. Pallett, “CSR-I (WSJ0) complete,” https://catalog.ldc.upenn.edu/LDC93S6A, 1993.
  48. [48]J. Engel et al., “Project Aria: A new tool for egocentric multi-modal AI research,” arXiv preprint arXiv:2308.13561, 2023.arxiv.org/abs/2308.13561
  49. [49]J.-M. Lemercier, J. Richter, S. Welker, and T. Gerkmann, “Analyzing diffusion-based generative approaches versus discriminative approaches for speech restoration,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023.arxiv.org/abs/2211.02397
  50. [50]W. Peebles and S. Xie, “Scalable diffusion models with transformers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 4195–4205.
  51. [51]“Method for the subjective assessment of intermediate quality level of audio systems,” International Telecommunication Union Radiocommunication Assembly, vol. 2, 2014.

Paper details

vlnc:W_u92vtz73
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

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

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents