Mesh2HRTF: Geometry-Driven HRTF Simulator

• Mesh2HRTF is a geometry-to-acoustics simulator that computes individualized HRTFs from 3D anatomical meshes using a robust boundary element method.
• Its numerical core, NumCalc, employs direct collocation with fast multipole acceleration to solve 3D frequency-domain acoustic boundary-value problems efficiently.
• The framework integrates into large-scale pipelines like SONICOM and WiDESPREaD, supporting both full-head and pinna-specific acoustics with reproducible mesh processing.

Mesh2HRTF is a geometry-driven, physics-based framework for the numerical synthesis of head-related transfer functions (HRTFs) from anatomical surface meshes. In the recent literature it appears both as an open-source package for listener-specific HRTF calculation and as the acoustic engine embedded in larger data-production pipelines, including the extended SONICOM HRTF dataset and the WiDESPREaD pinna-acoustics dataset. Its defining role is forward simulation: it transforms 3D geometry into HRTFs or HRIRs, typically in SOFA-compatible workflows, rather than estimating HRTFs from reverberant audio observations or reconstructing dense directional fields from sparse HRTF samples (Kreuzer et al., 2023, Poole et al., 7 Jul 2025, Guezenoc et al., 2020).

1. Conceptual scope and position within HRTF research

Mesh2HRTF occupies the forward-physics side of individualized spatial audio. Given a 3D mesh of the listener’s morphology and a specification of acoustic source or receiver positions, it computes individualized transfer functions by solving an acoustic boundary-value problem numerically. This differs fundamentally from audio-driven inverse estimation, sparse-grid interpolation, or continuous parametric representation. In that sense, Mesh2HRTF is not a generic HRTF personalization method; it is a geometry-to-acoustics simulator (Thuillier et al., 2024, Lee et al., 2022).

This distinction is important because adjacent literatures solve different problems. Score-based inverse estimation methods infer HRTFs from reverberant recordings and a learned prior rather than from anatomy (Thuillier et al., 2024). Learned interpolation methods estimate HRTFs at arbitrary target positions from neighboring HRTFs and anthropometric conditioning, which makes them suitable for densifying existing HRTF grids rather than replacing geometry-based simulation (Lee et al., 2022). Physics-informed neural upsampling reconstructs dense directional fields from sparse HRTF samples using Helmholtz-regularized neural networks, again complementing rather than supplanting a mesh-based solver (Ma et al., 2023). Hyperspherical-harmonic models provide a continuous post-processing representation over direction and frequency, but they presuppose discrete HRTF data and therefore sit naturally downstream of simulation or measurement (Szwajcowski, 2022).

A common misconception is that all individualized methods are interchangeable. The literature instead suggests a division of labor. Mesh2HRTF is the method of record when the objective is to propagate anatomical detail into physically interpretable HRTFs. Interpolation, inverse estimation, and continuous representations are best viewed as complementary layers around a forward solver, not as solver-level substitutes (Lee et al., 2022, Thuillier et al., 2024, Szwajcowski, 2022).

2. Numerical foundations and software architecture

The numerical core of Mesh2HRTF is NumCalc, which solves 3D frequency-domain acoustics via a direct collocation boundary element method with constant elements, stabilized for exterior problems by the Burton–Miller formulation and accelerated by single-level and multilevel fast multipole methods (Kreuzer et al., 2023). The governing exterior acoustic problem is written in boundary-integral form over a closed surface $\Gamma$ representing the scatterer. In the rigid-scatterer setting used throughout the HRTF literature surveyed here, the Neumann boundary condition is

$$\frac{\partial \phi(\mathbf{x})}{\partial n} = 0 \qquad \text{for } \mathbf{x}\in \Gamma.$$

A representative direct boundary-integral equation used in reciprocity-based HRTF computation is

$$\tau \phi(\mathbf{x}) = \int_{\Gamma} H(\mathbf{x},\mathbf{y}) \phi(\mathbf{y})\, d\Gamma(\mathbf{y}) - \int_{\Gamma} G(\mathbf{x},\mathbf{y}) v(\mathbf{y})\, d\Gamma(\mathbf{y}) + \phi_{inc}(\mathbf{x}),$$

with the usual interpretation of $G$ as the Helmholtz Green’s function and $H$ as its normal derivative (Ziegelwanger et al., 2016). After collocation discretization, the problem becomes a linear system. NumCalc supports prescribed pressure, prescribed normal velocity, and admittance boundary conditions, with rigid boundaries as the default and most relevant use case for Mesh2HRTF (Kreuzer et al., 2023).

Implementation details matter because Mesh2HRTF deliberately hides some numerical complexity from end users. NumCalc uses singular, hypersingular, quasi-singular, and regular quadrature treatments; a CGS iterative solver with row scaling or incomplete LU preconditioning; and automatic switching among conventional BEM, SLFMM, and MLFMM depending on problem size and wavelength-to-object-scale relations (Kreuzer et al., 2023). The paper on NumCalc is explicit that Mesh2HRTF users should inspect NC.out for the actual method used, cluster counts, expansion lengths, iteration counts, and residuals, because the solver does not validate mesh correctness in a comprehensive way (Kreuzer et al., 2023).

The practical implication is that Mesh2HRTF is not merely a file converter from mesh to SOFA. It is a wrapper around a substantial numerical acoustics stack whose reliability envelope depends on mesh quality, element regularity, local resolution, source placement, and solver convergence. The fast multipole machinery makes realistic HRTF-scale simulations feasible, but it does not relax the underlying BEM assumptions (Kreuzer et al., 2023).

3. Mesh preparation, anatomical representations, and grading

Mesh preparation is the decisive precondition for successful Mesh2HRTF use. The extended SONICOM pipeline states explicitly that Mesh2HRTF is based on the boundary element method and therefore requires a watertight surface mesh with no holes, no duplicate vertices, no self-intersections, and consistently outward-facing faces. Raw structured-light scans at approximately 0.5 mm resolution are therefore not passed directly to the solver; they are converted into simulation-compatible meshes with minimal filtering and smoothing and with interpolation of missing regions where necessary (Poole et al., 7 Jul 2025).

The SONICOM preprocessing sequence illustrates the practical anatomy-to-simulation transformation. After conversion to a watertight mesh, scans were aligned to the Frankfurt plane so that the face was consistently oriented forward. Head hair and facial hair were removed, and the geometry was truncated below the neck for consistency across subjects. Two processed variants were then created. The “pre-processed” scan is the minimally modified watertight version intended to preserve anatomy as much as possible. The “plugged” version additionally occludes the ear canal up to the entrance, reflecting the standard modeling assumption that the ear canal itself is not modeled as introducing direction-dependent effects and that the receiving point is effectively placed at the ear-canal entrance (Poole et al., 7 Jul 2025).

Large-scale Mesh2HRTF workflows rely heavily on nonuniform meshing. The extended SONICOM release used graded meshes inspired by curvature-adaptive mesh grading, keeping mesh density high around the ipsilateral pinna and reducing it on the contralateral side to lower numerical cost (Poole et al., 7 Jul 2025). This strategy descends directly from the a-priori mesh grading literature, which formalized the idea of prescribing a spatially varying target edge length as a function of normalized distance from the driven ear-canal region (Ziegelwanger et al., 2016). In that framework, the local target edge length is

$$\hat\ell_{\bar d_{\mathcal{E}}} = \hat\ell_{\min} + \left(\hat\ell_{\max} - \hat\ell_{\min}\right)\mu(\bar d_{\mathcal{E}}),$$

where $\mu$ is a grading function such as a power law or raised cosine and $\bar d_{\mathcal{E}}$ measures distance from the microphone or source region (Ziegelwanger et al., 2016).

The significance of grading is not merely computational. The a-priori mesh grading study showed that keeping the ipsilateral ear region fine and gradually coarsening toward the contralateral side can cut element count by roughly 7–10× while maintaining or improving numerical accuracy and predicted localization performance (Ziegelwanger et al., 2016). Later SONICOM-scale deployment confirms the same principle operationally, with separate left and right graded meshes released as PXXXX_graded_left.stl and PXXXX_graded_right.stl, implying ear-specific optimization consistent with the ipsilateral-resolution strategy (Poole et al., 7 Jul 2025).

4. Simulation workflows and released artifacts

In the extended SONICOM HRTF dataset, Mesh2HRTF is the simulation engine embedded into a public data-production pipeline rather than the methodological novelty itself. The dataset contains 300 measured HRTF sets and 200 synthetic HRTF sets, with the synthetic subset restricted to participants whose scans were of sufficient quality for processing and simulation. A key failure mode was excessive distortions around the pinnae, which made some subjects unusable for simulation (Poole et al., 7 Jul 2025).

The reported simulation setup is concise but highly structured. HRTFs were simulated from 0 Hz to 24 kHz in increments of 150 Hz, that is, on the discrete grid

$f_k = 150k \ \text{Hz}, \qquad k = 0,1,\dots,160.$

Source directions followed the SONICOM measurement grid so that synthesized data could be compared directly to the measured dataset at matched positions. The grid contains 793 source positions, spanning elevations from $-45^\circ$ to $$\frac{\partial \phi(\mathbf{x})}{\partial n} = 0 \qquad \text{for } \mathbf{x}\in \Gamma.$$0 and azimuth covering the full $$\frac{\partial \phi(\mathbf{x})}{\partial n} = 0 \qquad \text{for } \mathbf{x}\in \Gamma.$$1 around the listener in $$\frac{\partial \phi(\mathbf{x})}{\partial n} = 0 \qquad \text{for } \mathbf{x}\in \Gamma.$$2 increments (Poole et al., 7 Jul 2025). The released simulation products are HRIRs in SOFA format—HRIR_SONICOM_44100.sofa and HRIR_SONICOM_48000.sofa—together with the intermediate geometry files used to derive them (Poole et al., 7 Jul 2025).

A more specialized workflow appears in WiDESPREaD, where Mesh2HRTF is applied not to full head-and-torso meshes but to registered left-ear meshes closed with a cylindrical base, yielding pinna-related transfer functions rather than full HRTFs (Guezenoc et al., 2020). In that setting, the source is placed at the entry of the filled ear canal by reciprocity, the surrounding receiver grid is a 2-meter-radius icosahedral geodesic polyhedron with 2562 directions, and the frequency grid spans 0.1 kHz to 16 kHz in 100 Hz steps (Guezenoc et al., 2020). This workflow demonstrates that Mesh2HRTF is not confined to full-listener HRTF generation; it can also be used as a large-batch simulator for ear-only acoustics when the research target is monaural pinna structure rather than binaural whole-head scattering.

What these pipelines share is reproducibility of intermediate representations. The SONICOM release externalizes minimally processed watertight meshes, plugged ear-canal variants, and graded ear-specific meshes (Poole et al., 7 Jul 2025). WiDESPREaD externalizes a paired geometry-acoustics dataset built from registered ear shapes and FM-BEM simulations (Guezenoc et al., 2020). In both cases, Mesh2HRTF functions as a reproducible bridge from anatomy to acoustics.

5. Validation, perceptual evidence, and recognized limitations

The literature draws a sharp distinction between operationalizing Mesh2HRTF at scale and validating the perceptual sufficiency of its outputs. The extended SONICOM paper is deliberately modest on this point: it does not report a systematic objective comparison between measured and Mesh2HRTF-synthesized HRTFs, and it does not provide tables of spectral distortion, localization-model scores, correlation metrics, notch-frequency errors, ILD/ITD errors, or perceptual listening-test outcomes. The release is positioned instead as infrastructure for such evaluations, with future work explicitly reserved for testing the perceptual efficacy of the synthetic HRTFs relative to measured HRTFs (Poole et al., 7 Jul 2025).

Dedicated validation studies provide the missing perceptual evidence. A within-subject VR localization comparison across five levels of HRTF individualisation showed that high-resolution scan-based synthetic HRTFs, explicitly synthesized using Mesh2HRTF, matched individually measured performance on all tested polar metrics, while photogrammetry-based synthetic HRTFs showed the greatest degradation alongside KEMAR (Pirard et al., 29 Jun 2026). The central result is not simply that synthetic HRTFs can work, but that not all geometry-derived HRTFs are equal: high-resolution structured-light scans at 0.5 mm resolution yielded synthetic HRTFs that were behaviorally indistinguishable from individually measured HRTFs in that task, whereas lower-fidelity photogrammetry reconstructions did not (Pirard et al., 29 Jun 2026).

The focused photogrammetry evaluation sharpens this conclusion. Using 72-image photogrammetry captures for 150 SONICOM subjects and Mesh2HRTF synthesis on the resulting meshes, the study found that PR synthetic HRTFs preserved ITD cues but exhibited increased ILD and spectral errors. Behavioral localization showed that azimuth remained relatively robust, but median-plane localization collapsed: PR synthetic HRTFs were behaviorally worse than random measured HRTFs on great-circle error, polar accuracy, front-back confusion, and quadrant error (Pirard et al., 25 Mar 2026). The stated practical message is direct: current photogrammetry pipelines support individual HRTF synthesis but are limited by insufficient pinna morphology details and high-frequency spectral fidelity needed for accurate individual HRTFs containing monaural cues (Pirard et al., 25 Mar 2026).

This body of evidence also corrects a second misconception: “individualized-from-geometry” does not automatically imply perceptually adequate. The photogrammetry papers repeatedly attribute failure to insufficient preservation of pinna detail, and one of them states the key point in explicit terms: Mesh2HRTF is not the limiting factor here; mesh quality is (Pirard et al., 25 Mar 2026). For Mesh2HRTF, the dominant failure modes documented in current public studies are therefore geometric rather than solver-intrinsic.

6. Benchmarking resource, machine-learning substrate, and future interfaces

Mesh2HRTF has become a central substrate for public benchmarks rather than merely a standalone simulator. The extended SONICOM release is particularly strong because it combines measured HRTFs, processed 3D geometry suitable for numerical synthesis, Mesh2HRTF-generated synthetic HRIRs for a large subset, standardized directional sampling, and the Spatial Audio Metrics Toolbox for objective comparison (Poole et al., 7 Jul 2025). This paired-modality structure supports measured-versus-simulated analysis, alternate mesh cleanup strategies, different ear-canal treatments, and learned surrogates that approximate Mesh2HRTF outputs from geometry (Poole et al., 7 Jul 2025).

Machine-learning work already uses Mesh2HRTF in exactly this way. A graph-neural-network study on photogrammetry enhancement places Mesh2HRTF at two critical stages: first as the baseline simulator for raw photogrammetry meshes, and then as the acoustic validator for GNN-refined meshes. The central question is not visual upsampling per se, but whether geometric enhancement improves Mesh2HRTF-derived HRTF fidelity relative to high-resolution scan baselines and measured references (Pirard et al., 3 Oct 2025). The same logic underlies WiDESPREaD, where a statistical shape model of 119 left-ear scans is used to generate 1005 valid synthetic ears, each paired with an FM-BEM-simulated PRTF set (Guezenoc et al., 2020).

Other methodological families can be read as downstream or lateral interfaces to Mesh2HRTF rather than competitors. Learned interpolation can densify sparse simulated HRTF grids and bridge coordinate-system differences across datasets (Lee et al., 2022). Hyperspherical-harmonic models can compress discrete Mesh2HRTF outputs into a continuous joint space-frequency representation with substantially fewer coefficients than SH-per-frequency fitting (Szwajcowski, 2022). Physics-informed neural networks can upsample sparse Mesh2HRTF directional outputs using Helmholtz-regularized field reconstruction (Ma et al., 2023). Score-based inverse estimators suggest a plausible hybrid workflow in which Mesh2HRTF-generated HRTFs provide large-scale synthetic training data for priors used in reverberant audio inversion (Thuillier et al., 2024).

The resulting research picture is technically coherent. Mesh2HRTF remains the canonical forward simulator when individualized surface geometry is available and physically grounded synthesis is required. Around that core, current work is building preprocessing methods for accessible geometry acquisition, public benchmark datasets, analysis toolchains, learned geometric refinement, interpolation and continuous representations, and inverse estimators that may eventually use Mesh2HRTF outputs as priors or supervision. The trajectory suggested by the recent literature is therefore not the displacement of Mesh2HRTF, but its consolidation as the reference geometry-to-acoustics layer in individualized spatial audio research (Poole et al., 7 Jul 2025, Pirard et al., 3 Oct 2025).