Personalized Head-Related Transfer Functions

• Personalized HRTFs are tailored acoustic transfer functions that capture individual anatomical effects such as pinna, head, and torso filtering.
• Estimation methods include direct acoustic measurements, numerical simulations, anthropometric regression, and machine learning models for sparse-measurement upsampling.
• Their application enhances binaural synthesis and spatial audio perception in VR, gaming, telepresence, and hearing devices by significantly reducing localization errors.

Personalized head-related transfer functions (HRTFs) are subject-specific acoustic transfer functions that describe how sound from a spatial direction is filtered by the listener’s torso, head, pinnae, and ear canal before reaching the eardrums. In the time domain, the corresponding head-related impulse response (HRIR) $h(t,\mathbf{r},\mathbf{a})$ acts as a directional filter on the free-field pressure $p_0$, and the frequency-domain HRTF is its Fourier transform, $$H(\omega,\mathbf{r},\mathbf{a})=\mathcal{F}\{h(t,\mathbf{r},\mathbf{a})\}$$ (Sánchez et al., 6 Jan 2025). Because these transfer functions vary considerably with individual anatomy, personalized HRTFs are central to binaural synthesis, virtual and augmented reality, gaming, telepresence, hearing technologies, and auditory research; their estimation has consequently become a major topic spanning acoustic measurement, numerical simulation, anthropometric modeling, sparse-measurement upsampling, and generative learning (Guezenoc et al., 2020).

1. Acoustic basis and anatomical determinants

For a source in the far field at direction $\mathbf{r}=(\theta,\phi)$, the pressure at the eardrum can be modeled as a convolution between the free-field pressure and the HRIR, so HRTFs and HRIRs form a Fourier pair and can be treated either as complex transfer functions or as binaural impulse responses (Sánchez et al., 6 Jan 2025). In much of the recent machine-learning literature, HRTFs are also treated as spherical signals $H:\mathbb{S}^2\rightarrow\mathbb{R}^L$, indexed by direction on the sphere and by frequency bin, which makes spherical harmonic and neural-field representations natural (Chen et al., 2023).

Individuality arises from several anatomically distinct mechanisms. The pinna introduces direction-dependent spectral notches and peaks, especially above about $5$ kHz; the head causes shadowing and interaural level differences and contributes to interaural time differences; the torso and shoulders affect low- and mid-frequency coloration and elevation cues; and the ear canal introduces resonances (Sánchez et al., 6 Jan 2025). The survey literature treats these factors as the physical basis of HRTF individuality and links generic-HRTF failure modes directly to anatomical mismatch (Guezenoc et al., 2020).

Recent dataset construction makes this anatomy–acoustics relation explicit. The extended SONICOM dataset provides 300 measured subjects, 200 synthesized HRTFs generated using Mesh2HRTF, and pre-processed 3D scans of the head and ears optimized for HRTF synthesis; the scans are aligned to the Frankfurt plane, cleaned, truncated below the neck, and graded to concentrate mesh resolution near the ipsilateral pinna (Poole et al., 7 Jul 2025). This supports both data-driven and physics-based personalization pipelines in which morphology is no longer an abstract covariate but a geometric input.

2. Perceptual significance of personalization

The perceptual rationale for personalization is well established. Using non-individual HRTFs is associated with weak externalization, wrong perception of elevation, front–back inversions, and broader localization errors (Guezenoc et al., 2020). Recent application-oriented work adds reduced immersion, degraded speech understanding in multi-talker scenes, and poorer front–back and elevation discrimination to that list (Hu et al., 2 Oct 2025).

A particularly important spectral cue is the first notch frequency, $N1$, which is a dominant feature in elevation localization. The reported just-noticeable difference for $N1$ is $0.1$–$0.2$ octaves, making it both perceptually meaningful and practically useful for feature-level personalization (Arbel et al., 2024). This suggests that personalization need not always begin with full-spectrum reconstruction; accurate prediction of key individualized cues can already constrain HRTF selection or adaptation.

Objective and modeled-perceptual results support the same conclusion. In the diffusion-based HRIR predictor, the generated HRTFs achieve an absolute ITD error of $p_0$0, which is reported as below 1 JND in many conditions, while predicted and measured HRIRs align closely in onset timing and amplitude envelope (Sánchez et al., 6 Jan 2025). In the in-the-wild earbud-based system, personalized HRTFs derived from everyday recordings reduce front–back confusion from $p_0$1 with a generic HRTF to $p_0$2, approaching the $p_0$3 obtained with ground-truth anechoic HRTFs (Jayaram et al., 2023). A plausible implication is that magnitude personalization alone, even with generic phase, can recover a substantial part of the perceptually relevant individual structure.

3. Principal personalization paradigms

The classic survey organizes HRTF individualization into four families: acoustic measurement, numerical simulation, indirect individualization from morphology, and indirect individualization from perceptual feedback (Guezenoc et al., 2020). That taxonomy still describes the field, but each family has been extended by recent work.

Direct acoustic measurement remains the reference method. It typically requires loudspeaker arrays or moving loudspeakers, in-ear microphones, and controlled environments, and its main drawbacks are time, cost, and infrastructure (Sánchez et al., 6 Jan 2025). Recent work on continuous multi-channel measurement reformulates time-varying HRIR estimation as a state-space problem and uses a Kalman smoother with expectation maximization to learn the system model; in simulation, system distances are improved by up to $p_0$4 dB, and a quality level of $p_0$5 dB is reached at about $p_0$6 with the proposed method, versus about $p_0$7 for NLMS with white noise and about $p_0$8 for shifted NLMS with PSEQ (Kabzinski et al., 2021). This directly targets the measurement-speed bottleneck that constrains individualized datasets.

Numerical simulation uses 3D geometry and acoustic solvers rather than measured HRIRs. Mesh2HRTF, rigid-body Helmholtz modeling, and graded surface meshes now support large-scale synthesis of personalized HRTFs from scans (Poole et al., 7 Jul 2025). In parallel, deep geometry-to-HRTF prediction has become competitive: a CNN using spherical cap harmonic coefficients for pinna geometry and spherical harmonic coefficients for HRTF magnitudes and onsets reports global LSD values of $p_0$9 dB for the left ear and $$H(\omega,\mathbf{r},\mathbf{a})=\mathcal{F}\{h(t,\mathbf{r},\mathbf{a})\}$$0 dB for the right ear, and in the frontal direction its average LSD of $$H(\omega,\mathbf{r},\mathbf{a})=\mathcal{F}\{h(t,\mathbf{r},\mathbf{a})\}$$1 dB is lower than the $$H(\omega,\mathbf{r},\mathbf{a})=\mathcal{F}\{h(t,\mathbf{r},\mathbf{a})\}$$2 dB of the database-provided BEM simulation (Wang et al., 2022).

Morphology-based indirect methods range from database selection to regression. The survey notes that simple anthropometric selection and frequency scaling already improve localization relative to non-individual HRTFs, while more elaborate statistical regression remains promising but historically under-validated perceptually (Guezenoc et al., 2020). A recent image-based prototype operationalizes this route by detecting 55 ear landmarks, extracting 7 pinna measurements, and matching them to HUTUBS via Euclidean distance, thereby treating “ear image $$H(\omega,\mathbf{r},\mathbf{a})=\mathcal{F}\{h(t,\mathbf{r},\mathbf{a})\}$$3 anthropometry $$H(\omega,\mathbf{r},\mathbf{a})=\mathcal{F}\{h(t,\mathbf{r},\mathbf{a})\}$$4 database HRTF” as an end-to-end personalization pipeline (Pirard, 2023).

Perceptual-feedback approaches optimize HRTFs against what the listener hears rather than what a geometric model predicts. This line is now complemented by in-the-wild estimation from passive recordings: a UNet-based system uses only binaural recordings and head tracking, reaches a median LSD of $$H(\omega,\mathbf{r},\mathbf{a})=\mathcal{F}\{h(t,\mathbf{r},\mathbf{a})\}$$5 dB, and significantly improves localization and front–back confusion in listening tests without anechoic measurements or scans (Jayaram et al., 2023). This suggests a convergence between perceptual, behavioral, and data-driven personalization methods.

4. Representation learning and generative modeling

A major theme in recent work is the search for compact HRTF representations that preserve individualized structure while making prediction tractable. Standard low-dimensional spaces are usually optimized for spectral reconstruction, but not necessarily for perceptual compatibility. A recent perception-informed latent-space study shows that traditionally learned latent distances correlate only moderately with auditory metrics, then introduces Metric Multidimensional Scaling supervision and metric-based losses; on Sound Sphere 2, the proposed $$H(\omega,\mathbf{r},\mathbf{a})=\mathcal{F}\{h(t,\mathbf{r},\mathbf{a})\}$$6 training raises the correlation between latent distance and Predicted Binaural Coloration from $$H(\omega,\mathbf{r},\mathbf{a})=\mathcal{F}\{h(t,\mathbf{r},\mathbf{a})\}$$7 to $$H(\omega,\mathbf{r},\mathbf{a})=\mathcal{F}\{h(t,\mathbf{r},\mathbf{a})\}$$8 on test ground truth, while selection-based personalization reduces DRMSP from $$H(\omega,\mathbf{r},\mathbf{a})=\mathcal{F}\{h(t,\mathbf{r},\mathbf{a})\}$$9 to $\mathbf{r}=(\theta,\phi)$0 for the best candidate (Zhang et al., 3 Jul 2025). This indicates that a personalization space can be organized around perceptual geometry rather than only spectral MSE.

Cross-database learning introduces a different representational problem: measurement setups imprint database-specific spectral signatures. A normalization method based on per-position, per-ear average-person HRTFs substantially reduces those signatures, driving SVM database classification close to random guessing and improving cross-database reconstruction in HRTF field models; for example, in one experiment LSD drops from $\mathbf{r}=(\theta,\phi)$1 dB to $\mathbf{r}=(\theta,\phi)$2 dB (Wen et al., 2023). This is foundational for scalable personalization because pooled multi-database training is otherwise confounded by lab-specific transfer functions.

Generative models increasingly target HRIRs or compact filter parameterizations directly. A conditional DDPM trained on HUTUBS predicts time-domain HRIRs from 27 anthropometric features and DOA, uses 600 diffusion steps with a 1D U-Net denoiser, and reaches a global LSD of $\mathbf{r}=(\theta,\phi)$3 dB, within $\mathbf{r}=(\theta,\phi)$4 dB of a strong SH-based predictor, with an ITD error of $\mathbf{r}=(\theta,\phi)$5 (Sánchez et al., 6 Jan 2025). Neural IIR Filter Fields replace magnitude prediction by direct estimation of parametric IIR cascades and report better generalization to unseen directions under sparse measurements, especially when combined with low-rank adaptation (Masuyama et al., 2024). Retrieval-augmented neural fields go further by retrieving similar subjects from a dense dataset and conditioning the field on the retrieved HRTFs; with only 3 measurements, RANF reports ITD $\mathbf{r}=(\theta,\phi)$6, ILD $\mathbf{r}=(\theta,\phi)$7 dB, and LSD $\mathbf{r}=(\theta,\phi)$8 dB on SONICOM (Masuyama et al., 22 Jan 2025).

A plausible implication is that the field is moving away from a single dominant representation. Time-domain HRIR generation, SH-domain coefficient prediction, latent perceptual embeddings, and parametric IIR fields each capture different invariances, and recent results suggest that personalization quality is increasingly determined by whether those invariances align with the underlying anatomy, physics, and perceptual task.

5. Sparse-measurement upsampling and practical acquisition reduction

A large fraction of current personalization research targets the case where some individualized measurements are available, but only sparsely. This is distinct from zero-shot anthropometric prediction: the problem is to infer dense, listener-specific HRTFs from a few measured directions.

Spherical and transformer-based methods have substantially improved this regime. A spherical CNN that performs spherical harmonic transform, SH-domain convolution, and inverse transform reconstructs 480-direction HRTFs from 120 known directions on HUTUBS and reports an average LSD of about $\mathbf{r}=(\theta,\phi)$9, outperforming SH+DNN ($H:\mathbb{S}^2\rightarrow\mathbb{R}^L$0), HRTF Field ($H:\mathbb{S}^2\rightarrow\mathbb{R}^L$1), and CNN+GAN ($H:\mathbb{S}^2\rightarrow\mathbb{R}^L$2) while using 484,530 parameters (Chen et al., 2023). HRTFformer, a transformer operating in the SH domain with RoPE, token scaling, iterative projection, and neighbor dissimilarity loss, is explicitly designed for extreme sparsity; at sparsity 3 it reports polar accuracy error $H:\mathbb{S}^2\rightarrow\mathbb{R}^L$3 and quadrant error $H:\mathbb{S}^2\rightarrow\mathbb{R}^L$4, and at sparsity levels 3 and 5 it achieves the lowest ITD, ILD, and LSD among the compared methods (Hu et al., 2 Oct 2025).

Physics-informed models provide an alternative inductive bias. A PINN regularized by the Helmholtz equation uses sparse listener-specific measurements to interpolate and extrapolate HRTFs, with network width tied to spherical harmonic order; it outperforms SH and HRTF field baselines in high-frequency interpolation and in extrapolation to extreme elevations, particularly above $H:\mathbb{S}^2\rightarrow\mathbb{R}^L$5 kHz (Ma et al., 2023). This is especially relevant where direct measurements are incomplete because of rig geometry rather than by design.

Noise-robust sparse personalization is also receiving direct attention. A two-stage system combining an HRTF Denoisy U-Net with an AE-GAN upsamples from as few as 3 noisy measurements to 793 directions and reports LSD $H:\mathbb{S}^2\rightarrow\mathbb{R}^L$6 dB and cosine similarity loss $H:\mathbb{S}^2\rightarrow\mathbb{R}^L$7, with denoising outperforming wavelet filtering, spectral subtraction, and Kalman filtering on CSL, ILD, and ITD error (Hu et al., 24 Apr 2025). Spatial grouping offers yet another route: by partitioning directions into ipsilateral and contralateral regimes and using different grouping strategies on each side, a hybrid model reports mean LSD $H:\mathbb{S}^2\rightarrow\mathbb{R}^L$8 dB on seen angles and $H:\mathbb{S}^2\rightarrow\mathbb{R}^L$9 dB on unseen angles, outperforming both global and angle-specific baselines (Chang et al., 2024).

Taken together, these results suggest that “personalization from sparse measurements” is no longer a single method family but a design space organized by representation choice, spatial inductive bias, and measurement conditions. Some methods assume clean anechoic samples, others explicitly tolerate noise, and some are optimized for interpolation while others also target extrapolation.

6. Datasets, metrics, and unresolved issues

Progress in personalized HRTFs is inseparable from dataset design. HUTUBS provides 93 subjects with HRIRs and full anthropometric information and is widely used for leave-one-out evaluation (Sánchez et al., 6 Jan 2025). SONICOM now provides 300 measured subjects, 200 synthetic HRTFs, 793 directions, and 44.1/48 kHz SOFA files, together with the Spatial Audio Metrics Toolbox for ITD, ILD, and spectral-distortion analysis (Poole et al., 7 Jul 2025). Cross-database work shows, however, that simply increasing sample count is not enough if database-specific system responses remain entangled with anatomy (Wen et al., 2023).

Evaluation remains dominated by Log-Spectral Distortion,

$5$0

or closely related forms, alongside ITD and ILD errors (Sánchez et al., 6 Jan 2025). More recent work adds modeled perceptual measures such as Predicted Binaural Coloration, Auditory Externalization Perception, DRMSP, and localization-model outputs such as polar accuracy error and quadrant error (Zhang et al., 3 Jul 2025). This suggests a broadening consensus that spectral fidelity alone is insufficient, especially because localization-sensitive features and perceptual distance do not align perfectly with conventional reconstruction objectives.

Several unresolved issues recur across the literature. High-frequency detail above about $5$1 kHz remains difficult, especially when conditioning is limited to coarse anthropometry; the DDPM study reports that deviations mainly appear in that region, and its PBC analysis concentrates perceptual differences there (Sánchez et al., 6 Jan 2025). Robust handling of missing or noisy anthropometric measurements is typically not addressed explicitly (Sánchez et al., 6 Jan 2025). Phase and ITD personalization remain weaker than magnitude personalization in uncontrolled environments, which is why the in-the-wild system keeps generic phase (Jayaram et al., 2023). And even large new resources such as SONICOM still motivate further perceptual validation, more processed scans, and broader population coverage (Poole et al., 7 Jul 2025).

A final open question concerns the unit of personalization itself. Some methods personalize full HRTFs, some personalize latent coordinates, some personalize filter parameters, and some personalize only a dominant cue such as $5$2. The notch-frequency study shows that sub-JND N1 prediction requires roughly 200–300 training examples and that domain mixing helps only when feature distributions and acquisition modes are compatible (Arbel et al., 2024). This suggests that future systems may be hybrid: dense where data are available, feature-level where they are not, and perceptually constrained throughout.