When people label where a sound came from, they disagree, and the disagreement is not random. A compact microphone array spans a few centimetres, while speech and most everyday sounds below 1 kHz have wavelengths of 34 cm or more. At those frequencies the array can barely tell directions apart. Anyone listening to its recordings, and any model trained on them, hears a blurred scene. The practical claim of this note is that, for spatial audio, a large and computable share of human labelling disagreement is set by wave physics: the array geometry and the frequency content and loudness of each individual sound event. If that is true, a closed-form estimate of how uncertain each label should be can stand in for the multiple labels per clip that disagreement-aware training methods normally require. What the physics fails to explain is then a cleaner measurement of genuine human subjectivity.
This matters because spatial labels are expensive. Datasets such as STARSS23 rely on optical motion capture, meaning cameras that track infrared markers worn by participants. Human auditors then verify those tracks and settle disputes into a single answer, which discards the spread of their judgements. Methods that learn from annotator disagreement assume access to many raw votes per item. Beyond Black & White trains an auxiliary head on the empirical vote distribution, and Loss Modeling for Multi-Annotator Datasets uses per-annotator heads. Both require that many votes, both have only been tested on text, and both ignore where the disagreement comes from. We propose to measure human disagreement in direction-of-arrival labelling, show how much of it a physics bound predicts, and use the bound to train sound localizers from cheap single-annotator crowd labels.
Core Claim
For each sound event we compute a physics-predicted angular spread. For a source with per-frequency signal-to-noise ratio $\mathrm{SNR}(f)$, heard through a listening chain of effective aperture $D$ at speed of sound $c$, the standard small-aperture Cramér–Rao argument gives Fisher information for direction that grows with the squared phase difference across the aperture. The predicted spread is
$$\sigma_{\mathrm{phys}}^{-2} \;\propto\; \sum_{f} \mathrm{SNR}(f)\left(\frac{2\pi f D}{c}\right)^{2},$$
so an event whose energy sits at low frequencies or low SNR gets a wide $\sigma_{\mathrm{phys}}$. A loud, bright event gets a narrow one. The constant of proportionality is fitted once on held-out data.
We claim three things. First, $\sigma_{\mathrm{phys}}$ explains substantially more of the per-event variance in human direction labels than class identity alone. Second, if the aperture of the listening chain is changed experimentally, disagreement moves as the bound predicts. Third, training a localizer on single human labels, softened into targets of width $\sigma_{\mathrm{phys}}$ around each label, recovers most of the benefit of having many annotators per clip.
Direction-of-arrival errors inside the physics width are treated as expected blur. Errors far outside it are treated as mistakes or as real subjectivity. We expect the real subjectivity to concentrate on spatially extended or ambiguous sources such as running taps, doors, and music, and on events where the 360° video pulls listeners toward a visible object.
This differs from the closest work. PILOT predicts a learned Gaussian covariance per estimate, but that single covariance lumps all uncertainty sources together, and PILOT is trained on synthetic or motion-capture targets. Physically Informed Spatial Regularization for SELD penalizes predictions that are inconsistent with the array's spatial covariance, using an inverse Wishart model. It weights all frequencies uniformly and never involves human labels. Neither paper asks whether human disagreement is predictable from physics.
Research Program
Data collection. The first component is a re-annotation of part of the STARSS23 development set, since it is the only benchmark in our material that has independent motion-capture truth. Five crowd annotators per event will mark the direction of each labelled event segment, listening through head-tracked binaural rendering. Two variables are manipulated.
The first variable is the listening aperture. The same 32-channel spherical-array recording is rendered either at first order, which roughly matches the small tetrahedral array, or at fourth order, which is effectively a wider aperture. This makes the physics a controllable treatment rather than a correlate.
The second variable is whether the 360° video is shown. STARSS23 found that visual cues halve localization error for people but give no benefit for stationary or occluded sources. That suggests vision reshapes labels in a class-dependent way, which is a subjectivity effect worth isolating. It is analogous to the halo effect documented by Pérez Principi et al. in personality ratings.
Analysis. The second component models where the disagreement comes from. We regress each event's angular spread across annotators on $\sigma_{\mathrm{phys}}$, on class identity, and on both together. We then test whether the change in spread between first- and fourth-order rendering matches the change the bound predicts from the altered $D$.
The residual, meaning disagreement beyond the physics, is analysed separately. We borrow the loss-based idea from Loss Modeling for Multi-Annotator Datasets: errors separate into a coherent low-loss mode and a high-loss mode. Here we measure each label's distance from motion capture in units of $\sigma_{\mathrm{phys}}$, which should split careless errors from coherent minority readings of extended sources. That paper does this on text with learned losses. We do it with a physically grounded scale, so a large miss on a bass-heavy footstep is not mistaken for annotator error.
Training. The third component trains the localizer. Each human direction label becomes a von Mises–Fisher target, a Gaussian-like distribution on the sphere, with width $\sigma_{\mathrm{phys}}$. The localizer's direction head is trained against that target with the reverse (mode-seeking) KL divergence. We use that direction because Beyond Black & White found it the strongest auxiliary loss, which is the part of that paper's method we reuse.
The backbone is the audio-only STARSS23 baseline recurrent-convolutional network on the tetrahedral microphone format, with SALSA-Lite inputs from SALSA-Lite. All evaluation uses motion-capture truth, never human labels.
Baselines
Naive reference. The first reference point trains the backbone on raw single-annotator hard labels with no uncertainty handling. It measures how much damage crowd noise does.
Motion-capture upper bound. The second trains the same backbone on motion-capture labels. It defines the ceiling we are trying to approach cheaply. STARSS23 reports an audio-only tetrahedral localization error of 66.9° for its baseline, which is the reference for our backbone's scale. The strongest published system in the file, SMRL-SELD, reports a SELD score of 0.389 on STARSS23, against 0.461 for ADPIT. We do not expect to beat that architecture. Our comparisons are within a fixed backbone, and we will report where our best motion-capture model sits relative to 0.389 so readers can calibrate.
Full multi-annotator soft labels. The strongest competitor on the actual question is the approach of Beyond Black & White transplanted to directions. It uses all five annotators' labels per event as an empirical spherical distribution. It is strongest because it has direct access to the real disagreement, at five times the labelling cost. Its reported gains are text-only, from 85.73% to 86.76% accuracy on POS tagging and from 73.59% to 77.09% on stemming. We found no result on Lacuna for disagreement-aware training on spatial audio.
Learned-uncertainty baseline. A PILOT-style head learns its own per-event covariance from single crowd labels. We found no reported numbers on STARSS23 for it on Lacuna.
Ablations. The ablation that isolates the physics is a class-level smoothing baseline, which uses one fitted target width per class. A second ablation swaps $\sigma_{\mathrm{phys}}$ for a version computed with a wrong aperture $D$.
Fair comparison and success criterion. All systems share data splits, backbone, epochs, and a hyperparameter search of the same size. Width constants are tuned on a validation fold. Beating the field means two things. At equal labelling budget, the physics-width model must beat hard labels, learned covariance, and class-level smoothing on localization error and $F_{20^\circ}$, with non-overlapping bootstrap intervals, especially on low-frequency classes. At one-fifth the labels, it must match the five-annotator soft-label model within its confidence interval.
Why not just…?
Why not just average the annotators' directions and train on that? Averaging requires several labels per event and still collapses the spread into one point. The evidence from SONYC-UST is that consensus rules suppress faint sounds only a few listeners catch, and that paper had to adopt a "minority vote" rule instead. We include a mean-direction model trained on five labels as an extra reference. If it matches the physics-width model trained on one label, our cost argument still holds, but our accuracy argument does not.
Why not just let the network learn its own uncertainty, as PILOT does? This is a real threat. A heteroscedastic head, one that predicts its own noise level per input, might learn that low-frequency events are uncertain without being told. Our argument against it is identifiability. From one noisy label per event the network cannot tell physics blur from annotator error, so it can inflate uncertainty on events where annotators were simply careless. The first- versus fourth-order rendering experiment settles this. The physics bound predicts how disagreement should change when $D$ changes, whereas a learned head trained on the first-order data has no mechanism to transfer. If the learned head matches us on both renderings, we will report that physics adds interpretability but not accuracy.
Why not just smooth low-frequency classes more? This is the most dangerous alternative. If class identity already captures spectral content, a single width per class may explain disagreement as well as a per-event bound. The regression analysis and the class-level ablation test this directly. The per-event bound can only win through within-class variation in SNR, overlap, and loudness. If it does not add explained variance beyond class, the core claim fails and the note reduces to a dataset contribution.
Why not just discard low frequencies and localize from high-frequency cues? Speech and urban sound energy sits mostly below 1 kHz, according to the analysis in the hybrid post-filter paper. Phase-difference features also alias at high frequencies, and SALSA-Lite found that capping them near 2 kHz was more stable. We will still run a high-pass-input variant, so that the bound is not merely rewarding the model for ignoring information that the backbone cannot use.
Decisive Checks
The first check is whether, in the re-annotated STARSS23 subset, $\sigma_{\mathrm{phys}}$ adds significant explained variance in annotator spread beyond class identity. The second is whether moving from first- to fourth-order rendering narrows disagreement on low-frequency events by roughly the ratio the bound predicts, while leaving high-frequency events nearly unchanged.
The third check is whether, on motion-capture-scored test data, single-label training with physics-width targets closes most of the gap between hard crowd labels and motion-capture training. Beyond aggregate localization error, we will look for this gain on the low-spectral-centroid classes, footsteps, knocks, and doors, rather than on speech.
The fourth check examines the residual disagreement that physics leaves unexplained. We expect it to concentrate on extended sources and on audio-visual trials for visible people, and we will test whether it does. If it instead spreads evenly across classes, the separation of physics from subjectivity is not real, and we will say so.
References
- STARSS23: An Audio-Visual Dataset of Spatial Recordings of Real Scenes with Spatiotemporal Annotations of Sound Events
- Beyond Black & White: Leveraging Annotator Disagreement via Soft-Label Multi-Task Learning
- Loss Modeling for Multi-Annotator Datasets
- PILOT: Introducing Transformers for Probabilistic Sound Event Localization
- Physically Informed Spatial Regularization for Sound Event Localization and Detection
- On the Effect of Observed Subject Biases in Apparent Personality Analysis From Audio-Visual Signals
- SALSA-Lite: A Fast and Effective Feature for Polyphonic Sound Event Localization and Detection with Microphone Arrays
- Location-Oriented Sound Event Localization and Detection with Spatial Mapping and Regression Localization
- SONYC Urban Sound Tagging (SONYC-UST): A Multilabel Dataset from an Urban Acoustic Sensor Network
- A hybrid microphone array post-filter in a diffuse noise field