Extended SONICOM HRTF Dataset Overview

• The paper’s main contribution is providing an extended dataset that links measured and synthesized HRTFs, advancing machine learning for personalized spatial audio.
• It integrates data from 300 measured subjects and 200 synthesized HRTFs generated via Mesh2HRTF, supplemented with processed 3D scans and select demographic metadata.
• The inclusion of the Spatial Audio Metrics Toolbox standardizes evaluation through reproducible metrics, facilitating direct comparison across HRTF synthesis methods.

Searching arXiv for the dataset paper and closely related SONICOM papers to ground the article in current sources. The Extended SONICOM HRTF Dataset is a public resource for personalized spatial-audio research that enlarges the earlier SONICOM release by combining measured HRTFs, synthesized HRTFs derived from subject-specific geometry, pre-processed 3D scans, and a companion evaluation environment, the Spatial Audio Metrics (SAM) Toolbox. In its extended form, the dataset contains 300 measured subjects, synthesized HRTFs for 200 of those subjects generated with Mesh2HRTF, and demographic information for a consenting subset. The release is positioned as infrastructure for morphology-aware HRTF synthesis, machine learning, and standardized analysis, while preserving continuity with the previous SONICOM assets of measured HRTFs, unprocessed scans, and photogrammetry (Poole et al., 7 Jul 2025).

1. Expansion of the SONICOM resource

The extended release builds directly on the 2023 SONICOM dataset. The earlier release already combined measured HRTFs with photogrammetry and high-resolution 3D head scans for 200 listeners; the extended version increases the number of measured human subjects from 200 to 300 and adds a synthesized-HRTF component for 200 subjects whose scans were suitable for numerical processing. Demographic metadata are included for a subset of participants who consented to share them, and the metadata CSV contains age in years, ethnic group, and sex at birth (Poole et al., 7 Jul 2025).

The paper’s stated motivation is infrastructural rather than purely archival. Conventional HRTF measurement is described as expensive, slow, and error-prone, while public datasets remain too small and too heterogeneous for many machine-learning and personalization workflows. The extended release therefore aims to support scalable algorithm development by pairing measured acoustics with morphology and with numerically generated HRTFs from the same subjects. The larger subject count is presented as improving suitability for machine learning relative to the previous release, although the paper also notes only indirectly that 300 subjects remains modest relative to the dimensionality of individualized HRTF modeling.

Two common misconceptions are explicitly contradicted by the release description. First, the dataset does not provide synthetic HRTFs for all 300 measured listeners: the synthesized set covers 200 subjects only. Second, the demographic metadata do not establish population representativeness, because the paper does not provide a full demographic breakdown or a statistical representativeness analysis.

2. Data contents, directional organization, and file structure

The dataset is organized participant-by-participant, with each participant represented by a folder named PXXXX, where XXXX is the participant number. For subjects whose scans were suitable for simulation, an additional SYNTHETIC_HRTF subfolder is present. The release retains the previous SONICOM measured HRTFs, unprocessed 3D scans, and photogrammetry, and augments them with processed scan assets and synthesized SOFA files (Poole et al., 7 Jul 2025).

Asset	Scope	Notes
Measured HRTFs	300 subjects	Continuation of the SONICOM release
Synthesized HRTFs	200 of the 300 subjects	Generated using Mesh2HRTF
Processed 3D scans	Multiple mesh variants	Prepared for numerical acoustics and modification
Demographic metadata	Consenting subset	Age in years, ethnic group, sex at birth

HRTFs are organized on the SONICOM measurement grid of 793 source positions. The grid is described as spanning elevation from $-45^\circ$ to $225^\circ$, with azimuth sampled around the listener over $360^\circ$ in $5^\circ$ increments. In generic notation, the directional-frequency representation can be written as $H(\theta,\phi,f)$, with corresponding time-domain HRIRs $h(\theta,\phi,t)$. The synthetic outputs are released as SOFA HRIR files at 44.1 kHz and 48 kHz, so the resource can be used either in the time domain or, after FFT conversion, in the frequency domain.

For simulated subjects, the SYNTHETIC_HRTF subfolder contains the processed geometry and the derived binaural responses. The paper explicitly lists the graded simulation meshes as PXXXX_graded_left.stl and PXXXX_graded_right.stl, and the synthesized SOFA files as HRIR_SONICOM_44100.sofa and HRIR_SONICOM_48000.sofa. The release is publicly accessible from the Imperial College transfer portal specified in the paper and is distributed under the MIT license.

3. Geometry processing and synthetic-HRTF generation

The synthesized-HRTF component is generated with Mesh2HRTF, an open framework for numerical HRTF calculation based on the boundary element method. The source geometry originates from 3D scans captured as point clouds at approximately 0.5 mm resolution, including the head and parts of the shoulders. These raw scans are converted into watertight meshes using ExScan Pro Software. The paper states that this conversion is intentionally conservative: minimal filtering and smoothing are applied, with interpolation only where necessary to fill missing regions and satisfy watertightness requirements (Poole et al., 7 Jul 2025).

Mesh2HRTF imposes strict geometric constraints. The input mesh must be watertight, contain no holes, duplicate vertices, or self/intersecting faces, and have outward-facing normals. Several normalization steps are then applied. The face orientation is aligned to the Frankfurt plane, hair on the head and face is removed, and the scans are truncated below the neck. Two principal processed variants are produced. The “pre-processed” scan preserves anatomy with minimal modifications, while the “plugged” scan occludes the ear canal up to the canal entrance. The paper frames the plugged variant as consistent with a modeling convention in which the ear canal is not treated as contributing direction-dependent effects beyond the canal entrance.

To reduce computational cost, the meshes are graded according to curvature-adaptive principles. High resolution is retained around the ipsilateral pinna, while the contralateral side is downsampled. Separate graded meshes are provided for left- and right-ear simulation, and these graded meshes are the actual simulation inputs. Synthetic HRTFs are then simulated from 0 Hz to 24 kHz in increments of 150 Hz for every point on the 793-position SONICOM grid, i.e. over

$f \in \{0,150,300,\dots,24000\}\,\mathrm{Hz}.$

A notable design consequence is that the measured and synthetic data are spatially aligned on the same grid, which simplifies paired comparison and method development.

The paper is explicit that the synthesis contribution is infrastructural rather than a final validation of perceptual equivalence. It does not report a formal benchmark table or a detailed measured-versus-synthesized error analysis for the 200 synthetic subjects, and it does not present listening-test results for the synthetic HRTFs. Perceptual efficacy testing against measured HRTFs is framed as future work. This is an important boundary on interpretation: the synthetic HRTFs are released as research assets and validation targets, not as perceptually proven replacements for measured individualized HRTFs.

4. Spatial Audio Metrics Toolbox

The companion Spatial Audio Metrics Toolbox is introduced as the analysis layer of the release. SAM is described as a Python package, open source and under active development, intended to provide a standardized and reproducible framework for numerical evaluation, comparison, and visualization of HRTF data. The implemented metrics explicitly mentioned are spectral distortion and the binaural cues ITD and ILD (Poole et al., 7 Jul 2025).

The toolbox addresses a methodological problem rather than a data-collection problem. The paper argues that HRTF studies frequently rely on inconsistent metrics, custom scripts, and ad hoc visualizations, making cross-paper comparison difficult. SAM is therefore designed to support numerical and visual comparison of HRTFs from different subjects and synthesis methods, and to facilitate iterative improvement of both measurement and generation pipelines. The paper describes the package as robust and extensible, but it does not provide a package-level software specification, dependency list, or API-level details in the manuscript.

The paper also makes clear what SAM is not. It does not define a new mathematical benchmark metric in the text, and it does not instantiate a formal challenge protocol within the dataset paper itself. Its role is instead to regularize evaluation practice around a shared Python-based analysis environment. In that sense, SAM complements the dataset by reducing friction in comparing measured HRTFs, numerically synthesized HRTFs, and future derived reconstructions.

5. Position within the SONICOM research ecosystem

The extended release is closely connected to subsequent work that treats SONICOM as a benchmark substrate for reconstruction, upsampling, and geometry-driven synthesis. Sparse-to-dense HRTF magnitude upsampling on SONICOM has been studied with measurement regimes of 3, 5, 19, and 100 directions, where explicit frequency-domain modeling improves ILD and log-spectral distortion relative to earlier baselines (Chen et al., 12 Feb 2026). Retrieval-augmented personalization methods likewise use the 200-subject SONICOM release as a dense retrieval database for few-shot adaptation under the same sparse regimes (Masuyama et al., 22 Jan 2025). A time-domain alternative, BiFormer3D, treats SONICOM free-field HRIRs as the basis for grid-free sparse-to-dense binaural HRIR reconstruction (Xu et al., 30 Mar 2026).

The extended release also supports morphology-centered and accessibility-centered workflows. One line of work reconstructs low-resolution head meshes from SONICOM photogrammetry and trains a graph neural network to upscale those meshes for HRTF synthesis, using paired photogrammetry-derived and high-resolution scan geometry from SONICOM subjects (Pirard et al., 3 Oct 2025). Another study processes 72-image SONICOM photogrammetry captures for 150 subjects with Apple’s Object Capture API, synthesizes HRTFs from the resulting photogrammetry-reconstructed meshes, and compares them against measured HRTFs, scan-derived synthetic HRTFs, KEMAR, and random HRTFs, including a behavioural localization experiment with $N=27$ (Pirard et al., 25 Mar 2026).

Taken together, these studies place the extended dataset in a broader transition from static corpus to experimental platform. The measured cohort, paired synthetic data, and simulation-ready geometry make it possible to study sparse acquisition, mesh refinement, morphology modification, and synthesis validation within a common subject-linked resource. This suggests that the main significance of the extended release lies not only in enlarging SONICOM numerically, but in consolidating acoustics, geometry, and evaluation tooling into a single research substrate.

6. Limitations, interoperability, and open questions

Several limitations are explicit in the release description. Only 200 of the 300 measured subjects have synthetic HRTFs, because not all scans were of sufficient quality for numerical processing, with distortions around the pinnae identified as especially problematic. Demographic information is present only for a consenting subset, and the paper does not provide a demographic representativeness analysis. The measurement methodology of the measured HRTFs is not re-documented in the extended-release paper; readers are referred back to the original SONICOM publication for the full acquisition protocol (Poole et al., 7 Jul 2025).

The release also leaves open the question of interoperability with other HRTF corpora. A separate study on ten public HRTF databases shows that even when source positions are shared, database-specific measurement signatures remain strong enough that database identity can be classified unless a direction-dependent, ear-dependent normalization is applied; after normalization, classification falls to near random-guess level and cross-database reconstruction improves (Wen et al., 2023). Because SONICOM is not explicitly included in that study, transfer to Extended SONICOM is inferential rather than demonstrated. Still, this suggests that pooling Extended SONICOM with legacy databases should not be treated as a simple concatenation problem.

Spatial-grid heterogeneity is a second, distinct issue. Another line of work formulates global HRTF interpolation through local neighborhoods in Cartesian coordinates so that a single model can generalize across datasets with different spatial sampling distributions and coordinate conventions (Lee et al., 2022). Since the extended SONICOM release uses a fixed 793-position SONICOM grid while many other corpora use different layouts, the combination of cross-database normalization and grid-robust interpolation is a plausible requirement for multi-database learning, rather than a cosmetic preprocessing choice.

The most important unresolved point remains perceptual validation of the synthetic component. The extended-release paper explicitly stops short of claiming that its 200 synthesized HRTFs are perceptually interchangeable with measured HRTFs. Until such validation is supplied, the synthetic branch of the dataset is best understood as a large, paired simulation resource for benchmarking, morphology manipulation, and algorithm development, rather than as a finalized replacement for direct acoustic measurement.