---
title: 'SONICOM Dataset: Multimodal HRTF & Geometry'
url: https://www.emergentmind.com/topics/sonicom-dataset
type: topic
---

# SONICOM Dataset: Multimodal HRTF & Geometry

The SONICOM Dataset is a multimodal resource for individualized spatial audio research centered on head-related transfer functions (HRTFs), matched subject morphology, and synthesis-ready geometric assets. In its extended form, it provides 300 sets of measured HRTFs, 200 sets of numerically synthesised HRTFs generated with Mesh2HRTF, pre-processed 3D head and ear meshes, subject-level photogrammetry, and a companion Python analysis package, the Spatial Audio Metrics (SAM) Toolbox. The dataset is designed to support personalised HRTF synthesis, morphology-aware numerical acoustics, and machine-learning workflows that require matched acoustic and geometric data [2507.05053].

## 1. Release trajectory and dataset scope

SONICOM was introduced as a data-rich platform for immersive and personalised audio research and was later extended from 200 to 300 subjects with measured HRTFs. In the extended release, 200 of the 300 subjects also have synthesised HRTFs generated from subject-specific meshes using Mesh2HRTF, together with simulation-oriented mesh variants and associated metadata [2507.05053].

The extended release preserves the per-subject organisation used in the earlier dataset. Each subject is stored in a `PXXXX` folder containing measured HRTFs, 3D scans, and photogrammetry; for subjects with synthesis-ready scans, the extension adds a `SYNTHETIC_HRTF` subfolder containing meshes and synthetic HRTFs [2507.05053]. A demographic CSV is also provided for a subset of participants, linking participant IDs to age, ethnic group, and sex at birth [2507.05053].

The dataset’s scope is explicitly multimodal. Across the SONICOM ecosystem, the underlying subject records include acoustically measured HRTFs, 3D scans of heads and ears, photogrammetry inputs acquired with consumer hardware, and in the earlier SONICOM description also headphone transfer functions and depth pictures at different angles around the subjects’ heads [2510.02813].

## 2. Modalities and subject organisation

The extended dataset combines acoustic measurements, numerical simulations, morphology, and metadata within a consistent per-subject layout. Its core assets can be summarised as follows.

| Asset type | Coverage | Notes |
|---|---:|---|
| Measured HRTFs | 300 subjects | Stored per subject in SOFA-compatible SONICOM structure |
| Synthesised HRTFs | 200 subjects | Generated with Mesh2HRTF from subject-specific meshes |
| 3D morphology | 300 subjects; simulation-ready for 200 | Includes unprocessed scans and processed STL meshes |
| Photogrammetry | Per subject capture data | Used in downstream reconstruction workflows |
| Demographics | Subset of participants | Age, ethnic group, sex at birth |

The morphological component includes unprocessed 3D scans, pre-processed meshes, plugged meshes, and graded meshes. The photogrammetry component is sufficiently rich to support reconstruction experiments: SONICOM subjects were captured with 72 images per subject at 5° intervals around a full 360°, using an iPhone XS and Apple TrueDepth, with the captured data including high-resolution RGB images, depth maps, and gravity or IMU data [2603.24104].

This multimodality is not merely archival. Downstream work uses matched tuples of measured HRTFs, high-resolution EXScan Pro scans, and consumer-grade photogrammetry for the same individuals, enabling within-subject comparisons between measured acoustics, scan-derived numerical synthesis, and photogrammetry-derived numerical synthesis [2603.24104].

## 3. Measurement geometry, storage formats, and coordinate conventions

The extended SONICOM release provides dense spatial sampling. Synthetic HRTFs are computed on the same grid as the measurements: 793 positions, with elevation between \(-45^\circ\) and \(225^\circ\), and full azimuthal coverage from \(0^\circ\) to \(360^\circ\) in 5° steps [2507.05053]. This yields a high-resolution full-sphere sampling that supports directional analyses of binaural and monaural cues.

Measured HRTFs are acquired in an anechoic environment with a controlled loudspeaker arrangement, and simulation source positions are matched exactly to the SONICOM measurement grid [2603.24104]. The extended paper does not restate the full loudspeaker-array geometry, stimulus design, or head-stabilisation protocol; those details are inherited from the original SONICOM measurement setup rather than redefined in the extension [2507.05053].

The dataset uses SOFA as its principal exchange format for HRIRs. Synthetic HRIRs are distributed as `HRIR_SONICOM_44100.sofa` and `HRIR_SONICOM_48000.sofa`, with sampling rates of 44.1 kHz and 48 kHz, and with the usual SOFA dimensional structure \([M, R, N]\), where \(M\) is the number of measurements, \(R\) the number of receivers, and \(N\) the HRIR length in samples [2507.05053]. Measured HRIRs are likewise stored in SOFA, enabling aligned processing across measured and simulated conditions [2603.24104].

This storage design supports both direct perceptual use and algorithmic evaluation. A plausible implication is that SONICOM’s combination of dense directional sampling and standardized SOFA representation lowers the overhead for cross-method comparison, especially in studies that need strict directional correspondence between measured and synthesised responses.

## 4. Geometric assets and synthesis-ready mesh variants

A defining feature of SONICOM is the availability of high-quality 3D morphology explicitly prepared for BEM-based HRTF synthesis. Raw scans are point clouds with 0.5 mm resolution covering the head and part of the shoulders. These are processed in ExScan Pro with minimal filtering and smoothing, interpolation of missing regions for watertightness, Frankfurt-plane alignment, removal of head and facial hair, and truncation below the neck [2507.05053].

The extended dataset provides several mesh variants with distinct numerical roles.

| Mesh variant | File pattern | Purpose |
|---|---|---|
| Pre-processed mesh | `PXXXX_preprocessed.stl` | Watertight, anatomy-preserving baseline mesh |
| Plugged mesh | `PXXXX_plugged.stl` | Ear canal occluded up to the entrance |
| Graded meshes | `PXXXX_graded_left.stl`, `PXXXX_graded_right.stl` | Curvature-adaptive resolution for faster BEM |

The plugged meshes reflect the assumption that the ear canal itself does not provide direction-dependent filtering, with the reference point taken at the canal entrance [2507.05053]. The graded meshes follow a curvature-adaptive strategy in which resolution is kept high around the ipsilateral pinna and reduced farther away, yielding large reductions in BEM computation time while preserving accuracy near the ear [2507.05053].

Numerical HRTF synthesis uses Mesh2HRTF, which solves the exterior acoustic problem with a Boundary Element Method on a watertight head or head-torso mesh. In the extended release, the core simulation parameters are a frequency range from 0 Hz to 24 kHz, frequency spacing of 150 Hz, and evaluation at all 793 SONICOM directions [2507.05053]. In later SONICOM-based photogrammetry work, both scan-derived meshes and photogrammetry-derived meshes are post-processed so that measured, synthetic, and reference HRTFs can be compared on matched directions, aligned HRIR windows, and separated ITD metadata [2603.24104].

## 5. Tooling, benchmarks, and machine-learning usage

The SAM Toolbox is the canonical analysis companion to the extended SONICOM release. It is a Python package intended for efficient computation and visualisation of spatial audio metrics, including spectral distortion, ITD, and ILD, and is described as open-source and in continuous development [2507.05053]. The associated repository is `https://github.com/Katarina-Poole/Spatial-Audio-Metrics` [2507.05053].

SONICOM is explicitly positioned as a dataset for data-hungry methods. The extended release emphasizes rapid and iterative optimisation of HRTF synthesis algorithms, large-scale automatic generation of synthetic data, and improved suitability for machine learning because of the increased subject count and the availability of matched meshes and HRTFs [2507.05053]. The geometry is “morphology-ready” rather than “morphology-labelled”: the dataset provides usable high-quality surfaces, but anthropometric parameter sets or explicit shape annotations are not supplied in the extended paper [2507.05053].

One concrete example is sparse-measurement HRTF personalization. A retrieval-augmented neural field uses SONICOM as the only HRTF corpus for model development and evaluation, with 200 subjects in the base dataset, 793 measurements per subject, and 48 kHz sampling. After excluding subject P0079 because of atypical ITD behavior, the study uses 179 subjects for pre-training and 20 for held-out adaptation, evaluating personalization from 3, 5, 19, or 100 measured directions [2501.13017]. In that setting, SONICOM serves simultaneously as pre-training corpus, retrieval database, and evaluation benchmark, demonstrating that the dataset supports few-shot upsampling protocols closely aligned with the Listener Acoustic Personalization challenge [2501.13017].

These usages indicate that SONICOM occupies two roles at once: a reference corpus for controlled numerical acoustics and a substrate for learned personalization systems that exploit subject-to-subject variability in either geometry or acoustics.

## 6. Photogrammetry, reconstruction research, and empirical limitations

SONICOM is unusual in that it includes both high-resolution scan geometry and consumer-grade photogrammetry for the same subjects. In one SONICOM-based study, 72-image photogrammetry captures per subject were reconstructed with Apple’s Object Capture API and compared with measured HRTFs, scan-derived synthetic HRTFs, KEMAR, and random HRTFs across 150 subjects. The photogrammetry-reconstructed meshes captured global head and torso shape well, but the ear geometry was much coarser, with smoothed pinna features and reduced individualisation [2603.24104].

The empirical outcome of that comparison is sharply delimited. Photogrammetry-derived synthetic HRTFs preserved ITD cues but exhibited increased ILD and spectral errors. Across the 150-subject cohort, the reported median absolute ITD errors were approximately 21.8 µs for 3D synthetic HRTFs, 24.5 µs for PR synthetic HRTFs, 32.8 µs for KEMAR, and 34.7 µs for random SONICOM HRTFs. For ILD, the PR synthetic condition had the highest error at 2.8 dB, exceeding 3D synthetic at 1.7 dB, KEMAR at 1.8 dB, and random at 2.1 dB. For LSD, PR synthetic HRTFs again had the highest error at 10.2 dB, compared with 9.0 dB for 3D synthetic, 7.6 dB for KEMAR, and 6.6 dB for random SONICOM HRTFs [2603.24104].

The perceptual consequences were equally specific. In a behavioural localisation experiment with 27 SONICOM participants, azimuthal localisation was similar across conditions, but elevation and front-back performance strongly favored measured HRTFs. Reported great-circle errors were 23.4° for measured HRTFs, 30.5° for random SONICOM HRTFs, and 39.3° for PR synthetic HRTFs; quadrant errors were 15.4%, 27.1%, and 40.6%, respectively [2603.24104]. This supports the conclusion that consumer-grade photogrammetry as captured in SONICOM preserves head-size and ear-spacing information sufficiently for lateralisation, but not the fine pinna morphology required for high-frequency monaural cues.

A second SONICOM-based line of work uses paired PR meshes and EXScan Pro meshes to train a graph neural network for mesh upsampling. That study reconstructs 244 subject meshes from SONICOM photogrammetry using Apple Photogrammetry API, pairs each with a corresponding high-resolution EXScan Pro scan, computes bijective surface maps between the two, and trains a neural subdivision model with a Hausdorff Distance-based loss function. The upsampled meshes are then validated geometrically and through Mesh2HRTF, localisation, and Spatial Release from Masking experiments [2510.02813]. This suggests that SONICOM’s main long-term value may lie not only in direct use of the provided assets, but also in its capacity to supervise learned transformations from low-cost capture to synthesis-quality geometry.

## 7. Access, licensing, and position within the HRTF corpus landscape

The extended SONICOM dataset is publicly available at `https://transfer.ic.ac.uk:9090/#/2022_SONICOM-HRTF-DATASET/` and is released under the MIT license, permitting commercial and non-commercial use, modification, distribution, and private use subject to attribution and inclusion of the license text [2507.05053]. The data are distributed in standardized formats, principally SOFA for HRIRs and STL for mesh geometry, with a stable subject-centric folder structure [2507.05053].

Within the HRTF-dataset landscape, the extended SONICOM release is distinguished by the combination of scale, matched acoustics and geometry, and simulation-ready preprocessing. The extended paper contrasts SONICOM’s 300 measured subjects with smaller legacy corpora such as CIPIC with 45 subjects, Brinkmann’s cross-evaluated database with 96 subjects, Watanabe’s circular-array dataset with 105 subjects, and Yu’s near-field dataset with 56 subjects [2507.05053]. It also emphasizes that SONICOM provides high-resolution 3D head and ear scans for all 300 subjects and matched measured plus simulated HRTFs for 200 subjects, making it one of the largest datasets with paired measured/simulated HRTFs [2507.05053].

For high-fidelity psychoacoustics or perceptual evaluation, SONICOM’s measured HRTFs remain the reference condition in the downstream literature. For BEM benchmarking and morphology-aware synthesis, its EXScan Pro-derived meshes and simulation-ready variants are the primary assets. For large-scale personalization research, the coexistence of measured HRTFs, synthesis-ready geometry, and photogrammetry makes SONICOM a rare tri-modal substrate for evaluating the full chain from acquisition modality to acoustic prediction and perceptual outcome [2603.24104].

Source: https://www.emergentmind.com/topics/sonicom-dataset