HRTF Individualization

• HRTF individualization is the process of estimating subject-specific binaural transfer functions using anatomical details, sparse acoustic data, and deep learning methods.
• It employs methods such as acoustic measurement, geometry/anthropometry driven prediction, and latent techniques to capture both spectral and temporal cues.
• Recent advances integrate physics-based representations and hybrid sparse-measurement strategies to enhance perceptual accuracy while mitigating dataset biases.

Searching arXiv for recent and foundational papers on HRTF individualization to ground the article and citations. arxiv_search(query="HRTF individualization personalization scanned head geometry spherical harmonics anthropometric measurements deep learning", max_results=10) arXiv search: HRTF individualization personalization scanned head geometry spherical harmonics anthropometric measurements deep learning Head-related transfer function (HRTF) individualization is the estimation, selection, adaptation, or measurement of subject-specific binaural transfer functions that reflect the listener’s own pinnae, head, and torso, rather than a generic proxy. In contemporary work, the problem is framed not only as recovering direction-dependent magnitude and timing cues for all source directions, but also as reconciling practical constraints: exhaustive acoustic measurement remains burdensome, numerical simulation depends on high-fidelity geometry and substantial computation, and purely anthropometric predictors are limited by data scarcity and feature mismatch. Recent research therefore spans direct measurement, physics-based synthesis, anthropometry- and geometry-driven prediction, sparse-measurement correction, dataset normalization, and perceptually motivated selection or latent-space methods (Guezenoc et al., 2020).

1. Conceptual scope and methodological families

The central object is the directional, ear-specific transfer function $H(f,\Omega)$, or its time-domain counterpart $h(t,\Omega)$, whose individuality is perceptually consequential for localization accuracy, externalization, timbre, and front–back or elevation judgments. The survey literature distinguishes four broad routes to individualization: acoustic measurement, numerical simulation, indirect methods based on anthropometry, and indirect methods based on perceptual feedback. Acoustic measurement remains the perceptual reference but is equipment-intensive; numerical simulation is mobile and repeatable but depends on accurate geometry and still requires substantial processing; anthropometric and perceptual methods reduce acquisition burden but historically offered uneven perceptual validation (Guezenoc et al., 2020).

Family	Typical inputs	Representative work
Acoustic measurement and continuous acquisition	In-ear microphones, loudspeakers, sweeps or noise, head tracking	(Kabzinski et al., 2021, Tan et al., 2023)
Geometry- or anthropometry-driven prediction	3D head/ear scans, pinna measures, ear images	(Wang et al., 2022, Zhang et al., 2019, Arbel et al., 2024, Pirard, 2023)
Sparse-measurement hybrid correction	A generic or predicted HRTF plus a few subject samples	(Zandi et al., 2022, Thuillier et al., 2023, Hu et al., 12 Nov 2025)
Selection, upsampling, and latent matching	Sparse directional grids, candidate HRTF sets, latent embeddings	(Goldring et al., 2024, Hogg et al., 2023, Zhang et al., 3 Jul 2025)
Cross-database representation learning	Multiple HRTF databases with different measurement systems and grids	(Wen et al., 2023, Niu et al., 22 Aug 2025)

A persistent theme is that the “full” problem is global: a useful individualized HRTF set should cover all source-listener directions and preserve both monaural spectral structure and binaural timing/level cues. This is explicit in methods that model global HRTF fields over the sphere, rather than predicting direction-by-direction spectra (Wang et al., 2022).

2. Representations, parameterizations, and dataset unification

A major line of work reduces HRTF individualization to learning in compact, structured representations. One recent formulation represents global HRTF magnitudes and onsets with real spherical harmonics (SH),

$$H(\Omega, f) \approx \sum_{l=0}^{L}\sum_{m=-l}^{l} c_{lm}(f)\,Y_{lm}(\Omega),$$

while encoding local ear geometry with spherical cap harmonics (SCH) on a pinna-centered cap, thereby matching a compact geometry description to a compact directional acoustics description (Wang et al., 2022). Earlier statistical approaches used PCA or spatial PCA. In spatial PCA, the log-magnitude residual is decomposed into spatial principal components $W_q(\theta,\varphi)$, frequency- and subject-dependent weights $d_q(f,s)$, and a spatial mean residual $H_{\mathrm{av}}(\theta,\varphi)$; this separates directional structure from subject-specific weighting and supports continuous-direction prediction (Zhang et al., 2019).

Latent-variable approaches pursue a similar decomposition in learned coordinates. A recent autoencoder-based method explicitly constructs a “spatially independent latent representation” by averaging position-conditioned latent vectors over source positions, so that subject identity is encoded in a prototype $z_s$, while source position is reintroduced in the decoder through conditioning on position and frequency. This makes joint training across CIPIC and HUTUBS possible despite their different source-position grids (Niu et al., 22 Aug 2025). A complementary line argues that standard latent spaces optimized for reconstruction are not necessarily aligned with perceptual distance, and proposes Metric Multidimensional Scaling supervision plus perceptual losses so that Euclidean latent distances better reflect metrics such as Predicted Binaural Coloration, Auditory Externalization Perception, or DRMSP (Zhang et al., 3 Jul 2025).

Dataset heterogeneity is itself a technical obstacle. Cross-database learning is undermined by measurement-induced spectral coloration, microphone differences, room and rig effects, and mismatched spatial sampling. A normalization strategy based on a per-database, per-direction, per-ear geometric-mean magnitude equalization,

$$\tilde{H}_{i,E}^{D}(\Omega,f)=H_{i,E}^{D}(\Omega,f)/G_{D,E}(\Omega,f),$$

was shown to reduce database classification accuracy to chance level and to improve unified HRTF-field learning across multiple public databases (Wen et al., 2023). This suggests that a substantial part of “dataset bias” in HRTF learning is not anatomical but measurement-system-specific.

3. Geometry- and anthropometry-driven prediction

The most explicit geometry-to-acoustics pipeline in the recent literature predicts a listener’s global HRTFs directly from scanned head geometry and macro anthropometric measurements. In that system, the pinna is conformally mapped to the unit sphere, cropped as a spherical cap with $\theta_c=30^\circ$, remeshed on $\theta_c=25^\circ$, and represented by SCH coefficients up to degree $h(t,\Omega)$0, yielding $h(t,\Omega)$1 coefficients per ear. HRTF magnitudes are modeled with SH order $h(t,\Omega)$2 at 41 log-spaced frequency bins, and onset fields with $h(t,\Omega)$3. Two CNNs then predict SH coefficients of magnitudes and onsets from SCH descriptors and anthropometrics. On HUTUBS subjects with valid meshes, the resulting global LSD was $h(t,\Omega)$4 dB for the left ear and $h(t,\Omega)$5 dB for the right, with frontal-direction LSD $h(t,\Omega)$6 dB versus $h(t,\Omega)$7 dB for database-provided BEM simulations; onset prediction errors averaged $h(t,\Omega)$8 on the left and $h(t,\Omega)$9 on the right (Wang et al., 2022). The same work combines predicted magnitude and onset as

$$H(\Omega, f) \approx \sum_{l=0}^{L}\sum_{m=-l}^{l} c_{lm}(f)\,Y_{lm}(\Omega),$$0

with minimum-phase reconstruction and pure-delay onset insertion.

Anthropometry-only individualization remains attractive because it avoids scanning, but its reliability depends strongly on feature quality and dataset scale. A recent study targeting the first notch frequency $$H(\Omega, f) \approx \sum_{l=0}^{L}\sum_{m=-l}^{l} c_{lm}(f)\,Y_{lm}(\Omega),$$1, a dominant elevation cue, used the nine CIPIC pinna features to predict $$H(\Omega, f) \approx \sum_{l=0}^{L}\sum_{m=-l}^{l} c_{lm}(f)\,Y_{lm}(\Omega),$$2 with a three-layer neural network. On CHEDAR, after filtering, 903 examples remained and the neural model achieved $$H(\Omega, f) \approx \sum_{l=0}^{L}\sum_{m=-l}^{l} c_{lm}(f)\,Y_{lm}(\Omega),$$3 Hz RMS error, or $$H(\Omega, f) \approx \sum_{l=0}^{L}\sum_{m=-l}^{l} c_{lm}(f)\,Y_{lm}(\Omega),$$4 octave, compared with $$H(\Omega, f) \approx \sum_{l=0}^{L}\sum_{m=-l}^{l} c_{lm}(f)\,Y_{lm}(\Omega),$$5 Hz for linear regression and $$H(\Omega, f) \approx \sum_{l=0}^{L}\sum_{m=-l}^{l} c_{lm}(f)\,Y_{lm}(\Omega),$$6 Hz for the naive predictor. However, on smaller measured datasets such as HUTUBS measured and M1, errors remained near or above the lower just-noticeable difference, and domain mixing helped only when acquisition modality and feature distributions were compatible (Arbel et al., 2024). This underscores that anthropometric regressors are limited not simply by model class but by the availability of large, compatible, measured datasets.

Older PCA-based individualization already anticipated this trade-off. One horizontal-plane CIPIC method modeled magnitude HRTFs with ten principal components explaining $$H(\Omega, f) \approx \sum_{l=0}^{L}\sum_{m=-l}^{l} c_{lm}(f)\,Y_{lm}(\Omega),$$7 variance and predicted the corresponding weights by multiple linear regression from eight anthropometric measurements. It reported a mean individualization error of $$H(\Omega, f) \approx \sum_{l=0}^{L}\sum_{m=-l}^{l} c_{lm}(f)\,Y_{lm}(\Omega),$$8, improving substantially over earlier minimum-phase HRIR individualization by the same authors (Hugeng et al., 2010). Spatial PCA extended this idea to arbitrary directions by learning deep networks for spatial principal components, per-frequency anthropometric weights, a spatial mean residual, and ITD, all from CIPIC. With $$H(\Omega, f) \approx \sum_{l=0}^{L}\sum_{m=-l}^{l} c_{lm}(f)\,Y_{lm}(\Omega),$$9 spatial principal components, cumulative variance exceeded $W_q(\theta,\varphi)$0, and subjective localization tests showed performance comparable to a PCA baseline in most conditions while supporting arbitrary-direction prediction (Zhang et al., 2019).

At the opposite end of the complexity spectrum, an image-based recommendation pipeline used a CNN to detect ear landmarks on single 2D ear images, converted seven HUTUBS-style pinna distances from pixels to centimeters, and retrieved the nearest anthropometric neighbor in HUTUBS to assign that subject’s SOFA HRTFs. That work emphasized process validity rather than acoustic accuracy, and explicitly did not report objective spectral distortion or subjective listening tests (Pirard, 2023).

4. Sparse-measurement and hybrid personalization

A large recent literature treats individualization as a hybrid problem: start from a generic, predicted, or latent model, then personalize with a small number of subject-specific observations. One home-oriented system used a compact conditional variational autoencoder trained on ITA HRTFs to generate full-sphere HRTFs from sparse measurements acquired with commercial off-the-shelf components. It reported HRTF prediction improvements of $W_q(\theta,\varphi)$1 and $W_q(\theta,\varphi)$2 on average versus a non-adapted baseline in leave-one-subject-out experiments, LSD reductions of $W_q(\theta,\varphi)$3 for front-only and $W_q(\theta,\varphi)$4 for full-sphere adaptation with about 70 measurement locations, and an azimuth-error reduction of about $W_q(\theta,\varphi)$5 in a localization model. In hearing tests, correct azimuth identification increased from $W_q(\theta,\varphi)$6 to $W_q(\theta,\varphi)$7, a $W_q(\theta,\varphi)$8 increase (Zandi et al., 2022).

A more explicitly uncertainty-aware approach is the spherical ConvCNP meta-learner for HRTF error interpolation. It operates on time-aligned complex residual spectra, uses spherical set convolutions and a spherical CNN, and predicts Gaussian distributions over corrections from a sparse context set. On HUTUBS simulated HRTFs, it achieved up to $W_q(\theta,\varphi)$9 dB lower average LRE than thin-plate spherical splines over $d_q(f,s)$0–$d_q(f,s)$1 kHz, up to $d_q(f,s)$2 dB improvement in the $d_q(f,s)$3–$d_q(f,s)$4 kHz contralateral region, and reduced the number of needed samples for $d_q(f,s)$5 dB average LRE from about 50 to about 28, while maintaining uncertainty calibration within about $d_q(f,s)$6 dB for most predictions (Thuillier et al., 2023).

Graph-based methods extend the sparse-measurement paradigm by explicitly modeling spatial correlation across directions. GraphNF-SCA separates HRTF personalization from HRTF upsampling: a subject graph over retrieved neighbors predicts per-direction HRTFs for an unseen subject, then a direction graph refines those predictions using spatial-correlation augmentation. On SONICOM under the LAP Challenge 2024 setting, with only 3 measured directions, GraphNF-SCA achieved LSD $d_q(f,s)$7 dB and ILD $d_q(f,s)$8 dB, improving over GraphNF $d_q(f,s)$9, RANF $H_{\mathrm{av}}(\theta,\varphi)$0, and several selection baselines; similar gains were reported on HUTUBS measured, HUTUBS simulated, and CIPIC (Hu et al., 12 Nov 2025).

An even more aggressive reduction in measurement burden appears in a time-domain TCN approach that predicts HRIRs for target azimuths from a single 0° seed HRIR on the horizontal plane. Although objective SD and SDR degraded on a new measured dataset relative to the training dataset, psychoacoustic localization showed no significant main effect of HRTF type between measured and generated HRTFs, indicating that time-domain cue preservation can sometimes maintain performance despite spectral mismatch (Kobayashi et al., 2023).

5. Selection, upsampling, and alternatives to full synthesis

Not all personalization requires generating a complete individualized HRTF de novo. One practical alternative is direction-dependent selection from a small HRTF pool. In a VR experiment with five candidate HRTFs selected for distinct first spectral notch frequencies, choosing the best HRTF per direction reduced median elevation error from $H_{\mathrm{av}}(\theta,\varphi)$1 to $H_{\mathrm{av}}(\theta,\varphi)$2 relative to the best single HRTF, with $H_{\mathrm{av}}(\theta,\varphi)$3, while azimuth improvement was smaller and statistically insignificant. No single HRTF dominated across all directions, which the authors interpret as evidence that different source directions may be better served by different spectral profiles even for the same listener (Goldring et al., 2024).

Another alternative is upsampling from sparse individual measurements. A GAN-based method using a gnomonic equiangular projection and projection-aware convolutions learns to reconstruct dense HRTFs from sparse directional samples. On ARI data, barycentric interpolation was best when the input grid was relatively dense, but when inputs were sparse—especially 20 or 5 measured positions—the SRGAN outperformed SH upsampling, barycentric interpolation, and non-individual selection. At 5$H_{\mathrm{av}}(\theta,\varphi)$41280 positions, mean LSD was $H_{\mathrm{av}}(\theta,\varphi)$5 dB for SRGAN, $H_{\mathrm{av}}(\theta,\varphi)$6 dB for barycentric interpolation, and $H_{\mathrm{av}}(\theta,\varphi)$7 dB for SH; model-based localization metrics also favored SRGAN in that extreme sparse regime (Hogg et al., 2023).

These methods blur the line between “selection” and “individualization.” A perception-informed latent space makes the connection explicit by treating personalization as nearest-neighbor retrieval or top-$H_{\mathrm{av}}(\theta,\varphi)$8 selection in a latent geometry aligned with perceptual metrics. In that setting, nearest-neighbor personalization improved PBC- and DRMSP-based outcomes relative to a reconstruction-only latent space, albeit with a slight SDE trade-off (Zhang et al., 3 Jul 2025). A plausible implication is that future recommendation systems may combine latent retrieval, sparse user feedback, and local synthesis rather than committing to a single pipeline.

6. Evaluation, perceptual findings, and unresolved issues

Objective evaluation in HRTF individualization is dominated by spectral metrics such as log-spectral distortion, but the literature repeatedly shows that spectral fidelity is not the sole determinant of perceptual success. The CNN–SCH/SH geometry pipeline reported better sagittal-plane localization metrics than BEM simulation, with measured HRTFs giving $H_{\mathrm{av}}(\theta,\varphi)$9 and $z_s$0, predicted HRTFs $z_s$1 and $z_s$2, and BEM $z_s$3 and $z_s$4 (Wang et al., 2022). The TCN study likewise found that degraded SD on a new dataset did not significantly worsen behavioral localization (Kobayashi et al., 2023). This suggests that cue-specific preservation, especially of ITD and dominant monaural features, can sometimes matter more than uniform spectral agreement.

Recent perceptual comparisons have also sharpened the distinction between lateral and polar performance. In a within-subject VR study comparing five HRTF conditions—individually measured, KEMAR, random non-individual measured, high-resolution scan-based synthetic, and photogrammetry-based synthetic—lateral localization metrics were largely insensitive to HRTF type, whereas polar-domain metrics and confusion rates were strongly HRTF-dependent. Random measured HRTFs outperformed KEMAR on several polar metrics, high-resolution synthetic HRTFs matched individually measured performance, and photogrammetry-based synthetic HRTFs, together with KEMAR, showed the greatest degradation (Pirard et al., 29 Jun 2026). This is consistent with the broader observation that head width and coarse binaural structure suffice for many lateral tasks, while pinna fidelity and mesh resolution dominate elevation and front–back behavior.

Head movement complicates the interpretation further. In an audio augmented reality task with real visual anchors, individualized HRTFs improved perceived realism in static listening but not localization; when head movements were allowed, the pattern reversed, with individualized HRTFs improving localization but not realism or externalization relative to KEMAR (Martin et al., 10 Oct 2025). This does not imply a contradiction. Rather, it indicates that dynamic binaural cues can amplify the benefit of individualized spectral structure for localization while plausibility judgments saturate under matched room and visual cues.

Several limitations recur across methods. Magnitude-only modeling remains common, with minimum-phase reconstruction and separately estimated ITD or onset delays used as a practical surrogate for full phase (Wang et al., 2022). Small measured datasets continue to constrain anthropometric learning (Arbel et al., 2024). Cross-database bias remains substantial without explicit normalization (Wen et al., 2023). Mesh fidelity, especially of the pinna, is decisive for scan-based synthesis (Pirard et al., 29 Jun 2026). Sparse-measurement methods still depend on calibration quality, head tracking, or active sampling policy (Zandi et al., 2022).

Current directions therefore converge on a few themes. One is tighter coupling of geometry and acoustics through physics-aligned representations such as SH and SCH (Wang et al., 2022). Another is the use of unified, normalized, multi-database training to escape per-dataset sample scarcity (Wen et al., 2023). A third is hybrid personalization: initialize from geometry, anthropometry, or latent retrieval, then refine with a few measurements or perceptual queries (Thuillier et al., 2023). Finally, several works imply that the target of optimization should increasingly be perceptual compatibility rather than reconstruction alone (Zhang et al., 3 Jul 2025). This suggests that future HRTF individualization will be evaluated less as a single regression problem than as a joint problem of anatomical inference, sparse calibration, domain alignment, and perceptually constrained rendering.