Microphone arrays locate sounds by measuring tiny arrival-time differences between microphones. The classic tool is phase-transform cross-correlation (GCC-PHAT). It discards the loudness of every frequency and keeps only the phase, because room reflections corrupt loudness while the geometry of the source lives in the phase. Steered over a grid of candidate directions, this produces a "power map" (SRP-PHAT) that lights up where a sound probably comes from. The map has two well-known weaknesses. Standard implementations keep only its single brightest point, so they cannot report two simultaneous sources. In reverberant rooms, reflections create secondary bright spots that can outshine the true source. Deep sound-localization networks use these correlations only as fixed input features to a regressor that outputs one direction per sound class.
This note proposes to treat the phase-transform map as the starting distribution of a generative model rather than as a feature. Flow matching trains a network to move samples along nearly straight paths from a simple starting distribution to the data. It needs few solver steps when the start is already close to the target. Our model draws initial direction samples from the per-frame SRP-PHAT map, so they begin on the true source, on a reflection, or on a spurious peak. A learned velocity field then carries each sample to a ground-truth source direction. Because most of the mass already sits near real sources, the paths should be short and straight, and a handful of network evaluations should suffice. Because the output is a set of samples rather than one regressed point, overlapping sources and genuinely ambiguous frames are represented directly.
Core Claim
A data-dependent physical prior can replace Gaussian noise at the start of a flow. Here the prior is built from the array's own phase-transform map for the current frame. It yields a localizer that, at equal network size and compute, matches the strongest discriminative system on single-source frames and beats it on frames with overlapping sources. It should need roughly an order of magnitude fewer solver steps than the same flow started from noise. The claim is falsified if a Gaussian-started flow with identical conditioning reaches the same accuracy at the same step count. In that case the prior adds nothing and the gain, if any, comes from conditioning alone.
The closest work is TSFlow. It replaced isotropic noise with fixed Gaussian-process priors for univariate forecasting and reported competitive accuracy in 4 to 16 network evaluations against up to 100 for a diffusion baseline. That paper motivates the mechanism: an informed prior shortens transport. Its priors, however, are hand-picked kernels shared across all examples. Ours is computed from each input by a physics-based estimator, so the start distribution changes frame by frame with the room and the sources. On the acoustics side, Cordourier et al. stacked GCC-PHAT vectors over time ("GCC-grams") into a convolutional-recurrent network and cut direction error on DCASE 2019 Task 3. That system is purely discriminative and outputs a single direction per class. Every audio flow model in the file, including FlowSep, works on single-channel spectrogram latents started from Gaussian noise. FlowSep shows that rectified flow is effective and fast for audio, but it never touches spatial phase.
Research Program
Prior construction. For each time frame we compute SRP-PHAT over a spherical grid, following the open-source implementation evaluated in Lebarbenchon et al.. We then turn the map into a probability distribution with a temperature softmax mixed with a small uniform floor:
$$p_0(d \mid x) = (1-\epsilon)\,\mathrm{softmax}_d\big(P(d; x)/T\big) + \epsilon\, u(d),$$
Here $d$ is a direction on the sphere, $x$ is the multichannel frame, $P$ is the steered phase-transform power, $T$ is a temperature, and $u$ is the uniform density. The floor $\epsilon$ guarantees that the prior covers the whole sphere, so a source missed entirely by the map can still be reached. Both $T$ and $\epsilon$ are tuned on validation data.
Coupling and paths. Within each frame we draw $N$ prior samples. We pair them with the active ground-truth source directions by minimum-cost assignment on angular distance, replicating targets in proportion to their count. This is a per-frame version of the mini-batch optimal-transport coupling used to keep paths from crossing in TSFlow and AlignFlow. Because directions live on a sphere, the "straight line" is the great-circle arc between a prior sample $d_0$ and its target $d_1$. The network $v_\theta(d_t, t, h)$ regresses the arc's constant tangent velocity, where $d_t$ is the point at fraction $t$ along the arc and $h$ is an encoder embedding of the same GCC-gram and spectrogram input that Cordourier et al. used. Training reuses the conditional flow matching machinery in TorchCFM. At test time we integrate a few Euler steps and cluster the transported samples. Each cluster becomes one detected source direction, and its spread gives an uncertainty estimate. Sound-class activity comes from a detection head shared with the baselines, so any gain in localization cannot be attributed to better detection. For STARSS23 we extend the state with log-distance, started from a broad prior. This tests whether a generative output can sidestep the direction-versus-distance trade-off reported by the reverberation-feature study, where distance cues helped distance error but disturbed direction estimates.
Two-tier temporal structure. Hierarchical music models show that predicting coarse structure first constrains fine-grained decisions. Zeng et al. decode bar-level key and meter before note tokens and correct local transcription errors that way. PIANOTREE VAE nests notes inside simultaneous-note groups inside a score summary. We borrow this as an architectural constraint rather than as a music grammar. A segment-level encoder summarizes several seconds into a "scene layout" embedding: roughly where sources sit and how many there are. This embedding conditions every frame-level flow in the segment. The rationale is that performers and most real sources move slowly relative to the frame rate. A layout shared across frames should therefore suppress frame-level jumps to reflection peaks. For music scenes we align segments to bars, using the score timing available in the rendered data. For general scenes we use fixed-length segments. Music is the application regime where this matters most, because an ensemble is a canonical case of many simultaneous, partly correlated sources at stable positions.
Reverberation regimes as a non-linear lens. Environmental modeling work treats responses as threshold-governed rather than additive. In Yan et al., for example, leaf senescence reverses direction once drought crosses a severity threshold, and a linear fit smooths that reversal away. We expect phase-transform localization to fail the same way: reliably until the direct-to-reverberant ratio drops past some point, then abruptly. If so, the prior's usefulness also collapses there, and transport length and required steps should jump. We will measure the mean transport arc length from prior samples to targets, and the step count needed for a fixed accuracy, as functions of estimated direct-to-reverberant ratio. We will fit segmented (piecewise) regressions alongside linear ones. A clear breakpoint would justify an adaptive policy that spends more solver steps, or raises $\epsilon$, only below the threshold. Absence of a breakpoint is itself a reportable finding.
Baselines
The naive reference point is classical SRP-PHAT with iterative peak picking. It takes the global maximum of the steered power map, suppresses a neighborhood around it, and repeats up to the number of detected active classes, with the suppression radius tuned on validation. This is exactly our prior evaluated with zero flow steps. It therefore measures how much the learned transport adds over the physics alone. On LOCATA, Lebarbenchon et al. report a success rate (error below 20° in both angles) of at least 92% across tasks and 99.9% on the single static source task. On single-source LOCATA, then, we should not expect to beat it by much. Beating it means matching it there while clearly exceeding it on multi-source frames.
The strongest existing method on the shared benchmark is the fused GCC-gram plus spectrogram convolutional-recurrent network of Cordourier et al.. On DCASE 2019 Task 3 it reports a direction error of 20.4°, F-score of 87.1%, and error rate of 0.20. The same paper reports 30.8° for spectrograms alone. It is the right competitor because it already exploits phase-transform information with a deep network and differs from us only in being a deterministic regressor. We will reimplement it with the same encoder, data splits, augmentation, training epochs, and a hyperparameter search of equal size to ours, and confirm we reproduce roughly 20° before comparing. Beating it means a lower direction error at equal or fewer forward passes of comparable cost on the full set, and a clear margin on frames with two or more active sources. On STARSS23 the reference figures are the audio-only tetrahedral-format localization error of 66.9° from the STARSS23 paper, and the lowest relative distance error of 0.262 with the stpACC feature from the reverberation-feature study. We will rerun both under our protocol rather than rely on the quoted numbers.
The key ablation is the same flow with the same encoder, coupling, and training budget, but started from a uniform distribution on the sphere instead of the phase-transform map. It isolates the prior. We sweep both variants over 1, 2, 4, 8, and 16 solver steps and compare accuracy-versus-steps curves. A second ablation removes the segment-level layout embedding to isolate the hierarchical tier. A third replaces optimal-transport pairing with random pairing, to confirm that straightness comes from the coupling as well as the prior.
Why not just…?
Why not just train the GCC-gram regressor with a multi-output head? A discriminative network that predicts one direction per class, or a few fixed slots, is cheaper and is a genuine threat on single-source frames. We expect it to tie us there. Its weakness is structural. Under squared-error loss, a frame where a reflection and the true source are both plausible is pulled toward their average, which may be neither. Two instances of the same class must also be assigned to slots arbitrarily. The decisive evidence is the overlap-stratified comparison in the experiments. If the regressor matches us on multi-source and high-reverberation frames, the generative output is not worth its cost, and we will say so.
Why not just start the flow from Gaussian or uniform noise and condition on the GCC features? This is the most important alternative. The encoder sees the same information, so in principle the velocity field can learn to jump directly to the right peak. TSFlow's evidence that informed priors cut steps comes from forecasting, not acoustics, so it does not settle the question here. The uniform-start ablation decides it. Our claim requires the prior-started flow to reach a given accuracy in markedly fewer steps, for example at 2 to 4 steps where the uniform start needs 16. If the curves coincide, the prior claim is wrong.
Why not just use classical multi-peak SRP-PHAT and skip learning? It is transparent, needs no training, and is already strong on LOCATA. Its failure mode is that reflection peaks in reverberant rooms can be brighter than secondary true sources, so peak picking returns them. That is exactly what the learned transport is meant to correct. The zero-step baseline measures how much accuracy the flow adds, broken down by direct-to-reverberant ratio. If the gain appears only in low-ratio rooms, the honest conclusion is that the method is a reverberation corrector rather than a general localizer.
Why not just use a diffusion model? Curved diffusion paths typically need 50 to 1000 steps (Nakkiran et al.), and FlowSep already showed rectified flow beating diffusion for audio at far fewer steps. We therefore do not consider diffusion the binding competitor. We will still include a matched diffusion run at 50 steps as a sanity check on the speed claim.
Proposed Experiments
The primary benchmark is DCASE 2019 Task 3 with its four-channel tetrahedral array. We report error rate, F-score, direction error, and frame recall exactly as in Cordourier et al., and add direction error and F-score restricted to frames with two or more active sources. Real-room generalization is tested on STARSS22 (Politis et al.) and STARSS23 in tetrahedral format. On STARSS23 we add relative distance error for the distance-extended model. A direction-error regression relative to the distance-free model would show that the trade-off persists. Single-source and moving-source behavior, including robustness to array motion, is checked on LOCATA Tasks 1, 3 and 5 with the 12-microphone robot head. That is the regime where SRP-PHAT is strongest and our claim is only parity.
For the music regime, the file lists no spatial ensemble benchmark. We will therefore build a test set. Singing from GTSinger and piano rendered from POP909 scores are placed at fixed stage positions and convolved with simulated room responses spanning a wide range of reverberation, onto the same tetrahedral geometry. Score timing gives bar boundaries for the two-tier model. Because this set is synthetic, results on it are evidence about mechanism, not about real concert halls.
The decisive checks are three. First, accuracy-versus-steps curves for the prior-started and uniform-started flows on DCASE 2019 Task 3. The proposal stands only if the prior-started curve reaches the fused-regressor's level in clearly fewer steps. Second, overlap-stratified direction error against the reimplemented fused regressor, where we predict our largest margin. Third, the segmented-regression analysis of transport length and required steps against direct-to-reverberant ratio. A sharp breakpoint would validate the threshold view of localization failure and motivate adaptive step allocation. We also report calibration: the fraction of true sources falling inside the predicted sample spread. A generative localizer is worthwhile only if it is accurate and honest about ambiguity.
References
- Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting
- GCC-PHAT Cross-Correlation Audio Features for Simultaneous Sound Event Localization and Detection (SELD) on Multiple Rooms
- FlowSep: Language-Queried Sound Separation with Rectified Flow Matching
- Evaluation of an open-source implementation of the SRP-PHAT algorithm within the 2018 LOCATA challenge.
- AlignFlow: Improving Flow-based Generative Models with Semi-Discrete Optimal Transport
- Reverberation-Based Features for Sound Event Localization and Detection With Distance Estimation
- End-to-End Real-World Polyphonic Piano Audio-to-Score Transcription with Hierarchical Decoding
- PIANOTREE VAE: Structured Representation Learning for Polyphonic Music
- Drivers of the pre-season drought thresholds triggering earlier autumn foliar senescence in the Northern Hemisphere
- STARSS23: An Audio-Visual Dataset of Spatial Recordings of Real Scenes with Spatiotemporal Annotations of Sound Events
- Step-by-Step Diffusion: An Elementary Tutorial
- STARSS22: A Dataset of Spatial Recordings of Real Scenes with Spatiotemporal Annotations of Sound Events