A mirror of the SONICOM HRTF dataset from the Audio Experience Design group (Dyson School of Design Engineering, Imperial College London): acoustically measured head-related transfer functions (HRTFs), 3D head and torso scans, photogrammetry and numerically simulated HRTFs for several hundred listeners, plus KEMAR reference measurements.

This repository is a mirror of the dataset hosted by Imperial College London (transfer.ic.ac.uk, folder 2022_SONICOM-HRTF-DATASET), distributed under its original licence, MIT (see Licence). The original authors' folder layout and READMEs are kept unchanged. This is not the official distribution; for the canonical source, updates and corrections, use the SONICOM project.

The Hub dataset viewer is disabled (viewer: false). The data is a folder tree of SOFA / WAV / MAT / STL files, not a table of rows.

What is in it

Path	Contents
`P0001` … `P0441`	One folder per listener (435 folders; ids 253, 258, 270, 272, 275 and 435 are absent). Each holds a 3D scan, HRTFs at three sample rates, headphone EQ, synthetic HRTF inputs and outputs, plots and a README.
`KEMAR/`	Measurements of a KEMAR mannequin in many configurations: GRAS or Knowles microphones, with or without ear simulator, large or small pinnae, head orientations of 45° and 90°, and headset-mounted measurements (Meta Quest 3 BTE, HTC VIVE Focus Vision).
`KEMAR_GRAS_EarSim_LargeEars/`	The KEMAR reference listed in the metadata file.
`free_field_measurements/`	Free-field impulse responses of each loudspeaker (96 kHz WAV, 14 measurement sessions) and the system-compensation EQ filters derived from them (`reference_eq_001` … `014`, MATLAB).
`microphone_response/`	Characterisation of the Knowles in-ear microphones against a calibrated NTi XL2 / MA220 reference (REW `.mdat`, MATLAB, sweep signal).
`external_contributions/`	Third-party additions: `Brinkmann2025_EarPlugging` (effect of ear-canal blocking on simulated HRTFs for P0007, CC BY 4.0) and `Meyer2026_HRTFs with HWDs` (HRTFs with head-worn devices for KEMAR and ten listeners, CC BY 4.0).
`metadata_and_readme/`	`metadata.csv` (per-subject demographics, data availability flags, free-field EQ file used, average RMS levels), the original README and an outlier list (`important_information/`).

Per-listener folder (`Pxxxx/`)

Pxxxx/
├── 3DSCAN/            Pxxxx.stl, Pxxxx_watertight.stl, Pxxxx_Project1.asc (point cloud)
├── PHOTOGRAMMETRY/    72 HEIC photos at 5° steps, depth maps (TIF), camera poses (TXT)
├── HRTF/
│   ├── HRTF/{44kHz,48kHz,96kHz}/   SOFA and .3dti-hrtf files
│   ├── HPEQ/{44kHz,48kHz,96kHz}/   Sennheiser HD 650 personal headphone EQ (.mat, .wav)
│   └── plots/                      HRIR, ITD and horizontal-plane magnitude plots (PNG)
├── SYNTHETIC_HRTF/    preprocessed / plugged / graded meshes, MESH2HRTF output, measured-vs-synthetic comparison SOFA
└── README.txt

Not every listener has every part. The HRTF, Photogrammetry, 3D Scans and Synthetic HRTF columns of metadata_and_readme/metadata.csv say which parts exist (for example P0008, P0011 and P0012 have no synthetic HRTF).

HRTF file variants

For each sample rate (44.1, 48 and 96 kHz) the following SOFA files are provided. Each also has a NoITD version, and the NoITD versions are also provided as .3dti-hrtf for the 3D Tune-In Toolkit.

Variant	Description
`Raw`	Measured HRIRs, 50 ms long, no fades
`Windowed`	Windowed to 5 ms with fade in and out
`FreeFieldComp`	Windowed, with linear-phase free-field (loudspeaker and microphone) compensation
`FreeFieldCompMinPhase`	Windowed, with minimum-phase free-field compensation
`*_NoITD`	Onsets time-aligned so interaural time differences are removed. The applied shift in samples is stored in the SOFA metadata.

If you do not know which to use, start with FreeFieldCompMinPhase_NoITD_44kHz.3dti-hrtf (3DTI) or FreeFieldComp_48kHz.sofa (general SOFA tools).

Measurement set-up (summary)

Loudspeaker ring with 23 elevations from -45° to 225°, and a turntable for azimuth, giving a dense spherical grid.
Knowles FG-23329-P07 microphones at the ear-canal entrance (blocked canal); their response is characterised in microphone_response/.
The measuring system was re-characterised when the room or hardware changed. The Free Field EQ File column of the metadata names the filter used for each listener.

See the per-folder READMEs for details.

Intended uses

Personalised binaural rendering and headphone spatial audio
HRTF individualisation: predicting HRTFs from anthropometry, scans or photos
Comparing measured and numerically simulated HRTFs (MESH2HRTF)
Perceptual and objective HRTF evaluation, spatial-audio and auditory-model research

Loading

The files are SOFA (HDF5 / netCDF4), so use a SOFA reader rather than datasets.load_dataset:

from huggingface_hub import hf_hub_download
import sofar  # pip install sofar

path = hf_hub_download(
    "sdkv2/sonicom-hrtf-dataset",
    "P0001/HRTF/HRTF/48kHz/P0001_FreeFieldComp_48kHz.sofa",
    repo_type="dataset",
)
sofa = sofar.read_sofa(path)
hrir = sofa.Data_IR            # (measurements, 2 ears, samples)
fs = sofa.Data_SamplingRate
positions = sofa.SourcePosition  # azimuth, elevation, radius

The full tree is large (the Raw 96 kHz files alone are ~50 MB per listener, and scans add 100 MB or more). Download selectively:

hf download sdkv2/sonicom-hrtf-dataset --repo-type dataset \
  --include "P0001/HRTF/HRTF/48kHz/*" "metadata_and_readme/*"

Known issues and caveats

Outliers. metadata_and_readme/important_information/ lists HRTFs flagged for measurement problems (ILD/ITD hemifield correlation, cue-balance anomaly, left-right level difference at 0°, 0° over 5 dB). Check it before using a listener for training or evaluation.
Ear-canal plugging in meshes. Brinkmann (2025) found the ear-canal blocking in some SONICOM meshes differs from what was intended, which affects simulated HRTFs above about 4 kHz by up to 10 dB (shown for P0007 only). Treat synthetic HRTFs with care. See external_contributions/Brinkmann2025_EarPlugging/.
Scan alignment and holes. Meshes are approximately aligned and only lightly patched. Fill holes before 3D printing, and re-align before comparing models or using the automatic ear index finder in MESH2HRTF. P0001 is noted as having a misaligned preprocessed and plugged mesh.
Microphone response. The Knowles microphones deviate from the reference by about -2.8 dB (left) and -5.0 dB (right) on average, with a standard deviation of 2.8 dB and 1.6 dB. This is characterised but not removed from the data by this mirror.
Missing demographics. Age, ethnic group and sex at birth are blank for many listeners.
The original READMEs are versioned separately (v0.1, 2022 and v0.2, 2024), so file lists may differ slightly from them.

Personal data

The dataset contains 3D scans and photographs of listeners' heads and ears, plus age, ethnic group and sex at birth. These are potentially identifying. Use only under the terms the participants agreed to and the original licence, do not attempt re-identification, and do not redistribute scans or photos beyond what the licence allows. This mirror is private.

Licence

The SONICOM HRTF dataset is released under the MIT licence. The extended-dataset paper states that "the database is publicly available under the MIT license" at the Imperial download server, and the project's download page describes it as "publicly available for research purposes" and asks users to reference both papers below. The paper itself is CC BY 4.0, but that licence covers the article, not the data.

Exceptions and gaps:

external_contributions/Brinkmann2025_EarPlugging states CC BY 4.0 in its own readme.
external_contributions/Meyer2026_HRTFs with HWDs is licensed CC BY 4.0 (per the repository owner; no licence statement was found in the folder itself).
No standalone licence file was found in the dataset, so the MIT statement rests on the paper and the project page. Confirm with the dataset authors before making this mirror public.
The project page says the RGB image data is available from the authors on request, so check that the photogrammetry photos are covered by the open terms before redistributing them.

Citation

The project asks that you reference both papers when using the dataset: the original JAES paper and the extended-dataset paper (arXiv:2507.05053, Forum Acusticum 2025).

@article{engel2023sonicom,
  title   = {The SONICOM HRTF Dataset},
  author  = {Engel, Isaac and Daugintis, Rapolas and Vicente, Thibault and Hogg, Aidan O. T. and Pauwels, Johan and Tournier, Arnaud J. and Picinali, Lorenzo},
  journal = {Journal of the Audio Engineering Society},
  volume  = {71},
  number  = {5},
  pages   = {241--253},
  year    = {2023}
}

@inproceedings{sonicom2025extended,
  title     = {The Extended {SONICOM} {HRTF} Dataset and Spatial Audio Metrics Toolbox},
  booktitle = {Forum Acusticum},
  year      = {2025},
  note      = {arXiv:2507.05053}
}

Also credit the Audio Experience Design group (Imperial College London), F. Brinkmann for the ear-plugging study, and the authors of the Meyer 2026 head-worn-device HRTFs when you use those parts. Check the bibliographic details against the original source before publishing.