Spatial audio sounds convincing only when it is filtered through the listener's own head-related transfer function (HRTF). An HRTF is the direction-dependent filter that a person's head, torso and outer ear apply to incoming sound. Measuring one takes an anechoic chamber and a subject who sits still. Two cheaper routes exist. One is to choose a ready-made HRTF from a database. The other is to fit a model from a handful of measurements. Neither route answers a basic design question: how large must an HRTF catalog be before most new listeners will find an entry that sounds as good as their own? Microbial ecologists solve a structurally identical problem every day. They group noisy DNA reads into "operational taxonomic units" (OTUs), which are bins of sequences within a fixed similarity threshold. From how often bins are seen once or twice, they estimate how many units exist that they have never sampled, and what fraction of the next sample their catalog already covers.
We propose to transplant that machinery. Recorded human HRTFs, pooled across public databases, would be clustered into "acoustic OTUs" using a threshold set by perception rather than by spectral distance. Ecological richness and coverage estimators would then predict how much of the population a catalog of a given size serves. The claim can be checked directly: an estimate made from a small subsample of subjects should predict the fraction of held-out listeners who have a perceptually adequate match in a larger catalog. If the estimate holds, it gives HRTF database builders a stopping rule. It also gives personalization systems a principled triage: decide from a few measurements whether picking a catalog entry is enough, or whether a per-user model must be fitted.
Core Claim
The distribution of human HRTFs can be treated as a population of discrete perceptual types. Once subjects are binned at a threshold calibrated to localization performance, frequency-count estimators of unseen richness and sample coverage extrapolate catalog coverage to database sizes that no lab has yet measured. Existing work does not do this. The closest system is the retrieval-augmented neural field of Masuyama et al. (RANF). It retrieves the K nearest subjects by interaural time difference (ITD), meaning the arrival-time gap between the two ears. RANF then fuses those neighbours and adapts a neural field with low-rank updates, which won the 2024 Listener Acoustic Personalization (LAP) Challenge Task 2. RANF treats each stored subject as an isolated point. It has no notion of a bin, a threshold, or a population-level count. HRTF Field learns a continuous subject latent space across ten databases but never asks whether that space has discrete, countable structure. Our contribution is the counting and its validation, not a better upsampler.
Research Program
The first component is a common representation. Pooling databases is necessary because no single one is large: SONICOM has 200 subjects, HUTUBS 94 and CIPIC 45, and the ten databases unified by HRTF Field range from 187 to 2,818 directions per subject on incompatible grids. We use that paper's neural-field approach to resample every subject onto one shared direction set. A neural field here is a network that maps a direction to a filter value. HRTF Field models magnitude only, so we carry ITD separately, estimated from the measured impulse responses. Masuyama et al. showed that ITD is the more reliable signal of coarse head size, so it should not be discarded.
The second component is the distance, which is where the microbial analogy matters most. In ecology, the 97% identity rule is a convenient proxy for species. Its known failure is that it ignores real biological boundaries. This is the "arbitrary threshold" problem, and it is why strain-level studies such as Xue et al. tighten the threshold to 98.65%. The HRTF analogue of 97% identity would be log-spectral distance (LSD), the average gap between two magnitude spectra in decibels. HRTFformer shows that low LSD does not guarantee accurate localization. Zonooz et al. show that elevation depends on narrow, person-specific spectral notches in the 6–9 kHz band, which spectral averages blur.
We therefore define the distance between subjects $A$ and $B$ as $d(A,B)$. It is the increase in model-predicted localization error, specifically the polar accuracy and polar RMS error used by HRTFformer, when listener $A$ is rendered through $B$'s HRTF rather than their own. We symmetrize the distance by taking the larger of the two directions, $d(A,B)$ and $d(B,A)$. As a fallback we also use a notch-weighted spectral distance restricted to 4–12 kHz combined with ITD difference. The threshold $T$ is the degradation we accept. We sweep $T$ rather than fixing it, exactly as ecologists report results across thresholds.
The third component is the clustering itself. Greedy seed-based clustering, such as the CD-HIT approach used by Modha et al., is fast but depends on input order and suffers from chaining. Chaining is the effect whereby a run of intermediate items links dissimilar ones into one cluster. With only a few hundred subjects, scale is not the binding constraint, unlike the billions of proteins that forced metagRoot to use linear-time clustering. We can therefore afford medoid-based complete-linkage clustering. In that scheme every member of an OTU lies within $T$ of the representative and of each other. We report how stable OTU counts are across random input orders as a diagnostic.
Measurement noise between labs plays the role of sequencing error: it can inflate the apparent number of types. Ecology controls this with "mock communities," which are samples of known composition. Our analogue is any dummy head or repeated subject measured in more than one database or session, such as the KEMAR heads in CIPIC. If replicate measurements of one physical head fall into different OTUs at threshold $T$, that $T$ is below the noise floor and is discarded.
The fourth component is estimation. Let $N$ be the number of subjects, $S_{obs}$ the number of OTUs observed, and $f_1, f_2$ the number of OTUs containing exactly one and exactly two subjects. We compute the Chao1 richness estimate and the Good–Turing coverage estimate:
$$\hat S = S_{obs} + \frac{f_1^2}{2 f_2}, \qquad \hat C = 1 - \frac{f_1}{N}.$$
Here $\hat C$ is the estimated probability that a new listener falls into an already-observed OTU, and therefore has a catalog entry within $T$. Standard rarefaction and extrapolation curves built from the same counts predict coverage at catalog sizes $N' > N$.
The fifth component is the payoff for personalization. Given 3–5 measured directions plus ITD for a new user, we classify the user into an OTU or abstain, mirroring closed-reference OTU picking with open-reference fallback. If the user is confidently assigned, we serve the OTU medoid. Otherwise we route the user to a fitted model such as RANF.
Baselines
The naive reference point for selection is a generic HRTF. We use both the KEMAR dummy head and the population-mean HRTF, scored on the same held-out SONICOM test listeners and the same localization-model metrics. A random catalog subject is a second floor.
The strongest selection competitor is nearest-neighbour selection using the full measured HRTF. This is an oracle, because it sees data a real user would not provide. Thuillier et al. report that their nearest-neighbour oracle reaches a log relative error of −7.50 dB on HUTUBS, against −4.01 dB for a generic HRTF and −9.20 dB for their diffusion-prior estimate. This shows that catalog selection captures much of the available gain.
The strongest fitted-model competitor is RANF on SONICOM, trained with its published protocol and the same 179/20 subject split. Masuyama et al. report an ITD error of 20.5 µs at 3 measurements, versus 28.6 µs for a plain neural field, and an LSD of 3.0 dB at 100 measurements. HRTFformer reports an LSD of 4.20 dB, a polar accuracy error of 3.83°, and a polar RMS error of 35.37° at 3 directions on SONICOM. We do not expect medoid selection to beat these fitted models on average. Selection beats them only if, among users our triage marks as covered, the medoid's predicted localization error is within $T$ of RANF's at equal measurement count.
For the coverage claim itself, the baselines are estimators, not renderers. The naive estimator is the leave-one-out fraction of subjects who have some neighbour within $T$. The strongest alternative is a continuous density model: a Gaussian mixture or kernel density fitted to HRTF Field latent codes, with coverage computed as the probability mass within distance $T$ of stored subjects. All estimators receive the same subsample and the same distance. Our method wins if Chao1/Good–Turing extrapolation predicts held-out coverage at the larger catalog size with lower absolute error than both, averaged over resamples.
The key ablation swaps the perceptual distance for plain LSD at matched OTU counts. If LSD-based OTUs predict coverage and serve users equally well, the perceptual calibration is not doing the work.
Why not just…?
Why not just use RANF's K-nearest-neighbour retrieval, which already exploits similar subjects? RANF answers "what is the best estimate for this user?" It does not answer "how many users can a catalog of size N serve without fitting anything?" or "is this user covered?" Those are database-design and triage questions, and RANF has no mechanism to extrapolate them. We also test whether restricting RANF's retrieval pool to the user's assigned OTU changes its accuracy. If it does not, we report that OTUs add nothing to retrieval and confine the claim to counting.
Why not just compute leave-one-out nearest-neighbour coverage on the existing database? That number is an honest estimate of coverage at the current size, and it is essentially Good–Turing coverage in different notation. It cannot say what happens at 1,000 subjects. The decisive test is extrapolation: estimate from 50 SONICOM subjects, predict coverage at 179, and compare against the measured value. If leave-one-out with naive curve fitting extrapolates as well as Chao1-style estimators, the ecological machinery is unnecessary, and we will say so.
Why not just fit a continuous density on the latent space and skip discrete bins? This is the real threat, because HRTF variation may be genuinely continuous with no natural types. Frequency-count estimators can still work on arbitrary bins, but a density model might extrapolate better. We run it as a full baseline. We also check whether OTU counts plateau across thresholds as the catalog grows, which would suggest discrete structure. If the counts instead grow without bound in a way that tracks the volume of a continuous distribution, the result is still informative: it would show that no finite catalog suffices at strict thresholds.
Why not just measure more directions per user? Measurement is the bottleneck the field is trying to remove, as Thuillier et al. and Oden et al. document for children and other uncooperative subjects. Triage reduces measurement only for the users who need it.
Decisive Experiments
The central experiment uses SONICOM's 179 training subjects. We repeatedly subsample 25, 50 and 100 subjects, cluster at each threshold, and predict coverage of the 20 test subjects at catalog size 179. We score absolute prediction error against the empirical fraction of test subjects with a training subject within $T$. We then repeat on the pooled multi-database catalog, harmonized through HRTF Field, with whole databases held out. This tests whether a catalog built from some labs covers listeners measured in another, the analogue of cross-site generalization. Replicate dummy-head measurements set the minimum admissible threshold.
The second experiment tests triage. We measure OTU-assignment accuracy from 3, 5 and 19 SONICOM directions plus ITD, and the abstention rate at fixed precision. Among listeners classified as covered, we compare the medoid's predicted polar RMS error with RANF's at equal measurement count. Among listeners routed to RANF, we confirm that they are the ones selection would have failed.
The study succeeds if three things hold: extrapolated coverage beats both leave-one-out and density baselines, perceptual OTUs beat LSD OTUs in the ablation, and covered users are served by a single medoid at no meaningful perceptual cost. The negative outcome is equally reportable. If coverage keeps falling as the threshold tightens and no estimator extrapolates it, that would be evidence that per-user fitting cannot be avoided.
References
- Retrieval-Augmented Neural Field for HRTF Upsampling and Personalization
- HRTF Field: Unifying Measured HRTF Magnitude Representation with Neural Fields
- Comparison of Gut Microbiome Profile of Chickens Infected with Three Eimeria Species Reveals New Insights on Pathogenicity of Avian Coccidia
- HRTFformer: A Spatially-Aware Transformer for Personalized HRTF Upsampling in Immersive Audio Rendering
- Spectral Weighting Underlies Perceived Sound Elevation
- ViCTree: an automated framework for taxonomic classification from protein sequences.
- metagRoot: a comprehensive database of protein families associated with plant root microbiomes
- HRTF Estimation using a Score-based Prior
- A Dynamic Data Driven Computational Infrastructure for Reliable Computer Simulations