Latent Acoustic Mapping: Methods & Applications

• Latent Acoustic Mapping (LAM) is a framework that transforms acoustic signals into latent representations encoding spatial, physical, or task-specific structures.
• It leverages encoder-decoder architectures and self-supervised reconstruction to extract manifolds, improving analysis in scene mapping, source localization, and room characterization.
• LAM’s diverse applications—from UAV field modeling to inverse design and articulatory inversion—demonstrate its role in bridging measured data with physical models.

Latent Acoustic Mapping (LAM) denotes a family of methods that map acoustic observations, acoustic targets, or acoustically relevant metadata into a latent representation whose topology is intended to encode structure that is not explicit in the raw signal. In current literature, the term does not identify a single standardized architecture. Instead, it names several related constructions: spatio-temporal manifolds for in-the-wild acoustic scenes, latent physical variables for differentiable acoustic-field models, locally isometric spatial coordinates from multichannel transfer features, task-agnostic room embeddings, reciprocal latent fields for source–receiver propagation, articulatory regressors from acoustic embeddings, and latent control variables for inverse acoustic design (Montero-Ramírez et al., 2024, Iqbal et al., 2022, Cohen et al., 2023, Cámara et al., 2024, Götz et al., 2024, Seuté et al., 6 Feb 2026).

1. Terminological scope and representative usages

The literature uses “Latent Acoustic Mapping” for several technically distinct but structurally related mappings. In each case, acoustic information is not treated as an endpoint; it is transformed into a latent object that is subsequently interpreted, regularized, decoded, or optimized against a physical or task-specific constraint.

Setting	Latent object	Representative paper
In-the-wild scene analysis	VAE codes over event embeddings	(Montero-Ramírez et al., 2024)
UAV acoustic field modeling	latent source amplitudes for a monopole model	(Iqbal et al., 2022)
Acoustic scene mapping	2D locally isometric coordinates	(Cohen et al., 2023)
Acoustic-to-articulatory inversion	64-d acoustic latent projected to 6 PT controls	(Cámara et al., 2024)
Blind room-parameter estimation	task-agnostic AIR latent	(Götz et al., 2024)
Acousto-thermal holography	phase and amplitude masks on the hologram plane	(Cengiz et al., 2024)
Ventilated resonator inverse design	aligned geometry and response latents	(Cho et al., 2024)
DoA estimation	nonnegative spherical acoustic map	(Roman et al., 8 Jul 2025)
Precomputed sound propagation	reciprocal latent fields with symmetric decoders	(Seuté et al., 6 Feb 2026)
Audio-model security	Acoustic Latent Semantics from AGM priors	(Wang et al., 18 May 2026)

This breadth has two immediate consequences. First, LAM is best understood as a methodological pattern rather than a single benchmarked model family. Second, the latent variable can carry very different semantics: spatial position, source strength, room identity, articulation, geometry, phase mask, or cross-modal interference pattern. A common misconception is therefore to equate LAM exclusively with spatial localization or with VAE-based compression. The cited literature shows both restrictions to be too narrow (Montero-Ramírez et al., 2024, Cohen et al., 2023, Roman et al., 8 Jul 2025, Seuté et al., 6 Feb 2026).

2. Common formal structure

Across its variants, LAM repeatedly appears as a three-part construction: an encoder from observations or targets into a latent variable, a constraint or inductive bias that shapes the latent space, and a decoder or downstream operator that renders the latent useful. A recurring composition is

$z = T(x;\theta), \qquad \hat{p} = P(z;\phi),$

where $x$ is an acoustic observation or control input, $z$ is the latent variable, and $P$ is a task-specific physical or analytic decoder. In the physics-infused UAV formulation, $z$ is a vector of source amplitudes and $P$ is a differentiable interference model (Iqbal et al., 2022). In reciprocal latent fields, the pairwise predictor takes the form

$\hat{p}(a,b) = h\big(f(a), f(b)\big),$

with a symmetric decoder $h$ chosen specifically to enforce reciprocity (Seuté et al., 6 Feb 2026).

A second recurrent pattern is latent learning by self-supervised or unsupervised reconstruction. The in-the-wild acoustic-scene pipeline trains a per-user linear VAE on flattened event embeddings using the standard ELBO,

$$\mathcal{L}_{ELBO}(x) = \mathbb{E}_{q_\phi(z|x)}[\log p_\theta(x|z)] - KL\big(q_\phi(z|x)\,\|\,p(z)\big),$$

and then analyzes latent codes with cosine similarity and t-SNE (Montero-Ramírez et al., 2024). The ventilated-resonator AR-VAE splits the latent into a stochastic geometry code $z_p$ and deterministic alignment codes $x$0 and $x$1, with the total loss

$x$2

thereby using latent alignment to couple geometry and acoustic response (Cho et al., 2024).

A third pattern is analytic latent decoding through known acoustics. In self-supervised DoAE, the latent spherical acoustic map is constrained through covariance fitting,

$x$3

so that the learned map is interpretable as a spatial energy distribution over steering directions (Roman et al., 8 Jul 2025). This suggests that a defining feature of LAM is not merely dimensionality reduction, but the use of latent variables as an interface between data and structure: physical structure, spatial structure, geometric structure, or cross-modal structure.

3. Spatial and scene-centric mappings

One major usage of LAM concerns latent spaces whose neighborhood relations reflect where an acoustic observation occurred or which scene it belongs to. In “Spatio-temporal Latent Representations for the Analysis of Acoustic Scenes in-the-wild,” the pipeline is explicitly described as audio $x$4 detected acoustic events $x$5 vector embeddings (TF-IDF + Node2Vec) $x$6 VAE latent codes, yielding a latent manifold aligned to sparse, noisy location and situational labels (Montero-Ramírez et al., 2024). The WE-LIVE dataset contains 14 women volunteers with almost-continuous recording for 7–10 days each, with data from one participant discarded due to insufficient audio. Per user, the study uses approximately $x$7 hours of unlabeled audio and $x$8 hours of audio annotated with GPS cell location tags. Audio is normalized, converted to 16 kHz mono, segmented into 60 s clips, and processed by YAMNet into a $x$9 event-probability matrix per clip. Probabilities are thresholded at the user-specific 99th percentile, yielding approximately 5 triggered events per second, and the concatenated TF-IDF plus 5-dimensional Node2Vec representation gives a $z$0 embedding matrix per clip, flattened for VAE input. The reported qualitative result is that, before encoding, distributions appear random-like, whereas after VAE encoding the latent space shows clearer differentiation among acoustic scenes, notably between indoor locations and metro/subway clusters (Montero-Ramírez et al., 2024).

A more explicitly geometric formulation appears in LOCA-based acoustic scene mapping. There, the latent variable is a 2D coordinate system learned from 760-dimensional Relative Transfer Function features, with the constraint that local embedded covariances should be isotropic up to a known scale. The whitening loss,

$z$1

is combined with an autoencoder reconstruction loss, and the resulting embedding is aligned to ground-truth scan coordinates by Procrustes analysis (Cohen et al., 2023). In the reported simulations, LOCA achieves MAE values of 11.3 cm, 13.4 cm, and 18.5 cm at $z$2, 360, and 610 ms, respectively, compared with 72.2–74.3 cm for PCA and 13.7 cm degrading to 74.8 cm for the TDOA baseline under high reverberation (Cohen et al., 2023). The claim is therefore not simply that a latent space exists, but that the latent geometry is approximately isometric to the physical scan plane.

Self-supervised DoAE extends this scene-centric line to spherical acoustic maps inferred directly from multichannel covariance. LAM processes each of $z$3 frequency bands linearly spaced from 1.5–4.5 kHz, uses a learnable back-projection followed by four denoising 1D-convolution stages, and reconstructs the input cross-spectral matrix through the steering matrix (Roman et al., 8 Jul 2025). The base model has about 16K parameters per frequency band. On LOCATA, the 32-channel self-supervised variant LAM$z$4K-means reports 13.69° LE and 94.0 LR, while LAM$z$5GRU-MHSA reports 13.41° LE and 94.7 LR; on 4-channel LOCATA, UpLAM$z$6GRU-MHSA reports 14.44° LE and 84.26 LR, outperforming SELDnet and EINV2 on LE (Roman et al., 8 Jul 2025). Taken together, these studies show three distinct spatial interpretations of LAM: latent scene clustering, latent coordinate recovery, and latent acoustic imaging.

4. Physics-infused propagation and room-acoustic mappings

A second major lineage uses LAM to encode physically meaningful latent variables that parameterize an acoustic forward model. In OPTMA-Net for UAV acoustic fields, the transfer network maps a field point $z$7 to a latent vector $z$8 of monopole source amplitudes, and the physics layer computes

$z$9

The acoustic field model is written inside PyTorch so that reverse-mode auto-differentiation propagates through the physics layer (Iqbal et al., 2022). In the reported UAV experiment, data consist of approximately 1700 SPL samples around a hovering DJI Phantom, with a partial physics model based on $P$0 monopoles. Under a quadrant split, normalized MSE is 0.129 for the pure model, 0.089 for the sequential hybrid, and 0.034 for OPTMA-Net; under a radial split, values are 0.055, 0.041, and 0.013, respectively, corresponding to extrapolation performance that is approximately 3.8× to 4.2× better than the pure data-driven baseline (Iqbal et al., 2022). Here the latent variable is explicitly a physically interpretable pseudo-source description.

Blind room-parameter estimation adopts a different separation: the latent variable is first learned from Acoustic Impulse Responses and only then approximated from reverberant speech. The AIR-VAE stage models mel-spectrogram RIRs with

$P$1

and a separate speech encoder is trained to minimize $P$2 (Götz et al., 2024). The dataset contains 6,269 measured AIRs from ACE, EchoThief, IKS Aachen, OpenAIR, WDR, TU Ilmenau, MIT survey, and Aalto, with approximately 20 h of reverberant speech in 4 s segments. The downstream task is joint estimation of reverberation time $P$3 and clarity $P$4 in seven octave bands. At 1 kHz, the proposed method reports MAE 0.095 s and PCC 0.962 for $P$5, compared with 0.119 s and 0.947 for the end-to-end baseline and 0.186 s and 0.853 for the contrastive baseline (Götz et al., 2024). This formulation treats the latent space as “room-centric” and task-agnostic.

Reciprocal Latent Fields generalize the same idea to pairwise source–receiver propagation in large scenes. The encoder is a volumetric grid of trainable embeddings interpolated at a 3D position, and the decoder is symmetric, so that reciprocal parameters satisfy

$P$6

The paper studies Euclidean, Riemannian, and symmetric-MLP decoders, with the midpoint Riemannian approximation

$P$7

(Seuté et al., 6 Feb 2026). Reported uncompressed float32 memory is 3.1 GB for wave coding versus 1.8 MB for RLF in Audio Gym, and 182 GB versus 14 MB in Wwise Audio Lab. In a MUSHRA-like subjective listening test, mean scores are 61.8 ± 4.3 for RLF, 62.6 ± 4.4 for ground truth, and 27.7 ± 4.7 for the low anchor, with no significant difference between RLF and ground truth ($P$8) (Seuté et al., 6 Feb 2026). In this setting, LAM becomes a compact representation of a propagation field rather than of a signal class.

5. Inverse design, acousto-thermal control, and metamaterials

LAM is also used for inverse problems in which the latent variable is not inferred from an acoustic measurement, but from a desired thermal or acoustic outcome. In holographic thermal mapping with acoustic lenses, the latent acoustic space is the pair of aperture-field functions

$P$9

defined over the lens surface $z$0 (Cengiz et al., 2024). These latent functions are obtained by time reversal and used to reconstruct a target focused ultrasound field, which is then converted to heat through

$z$1

followed by the Pennes bioheat equation. The reported U-Net maps desired thermal patterns at the target layer to predicted hologram-plane latent functions $z$2 at $z$3 MHz, after which the design is converted into a phase-only hologram by BERD (Cengiz et al., 2024). In thin-sample experiments with 10 s continuous exposure, the measured temperature rise reaches up to approximately 3 °C, with PSNR 10.73–13.95 dB, RMSE 0.20–0.29 °C, simulated SSIM approximately 0.76, and measured SSIM approximately 0.32–0.33 (Cengiz et al., 2024). The latent variable here is a manufacturable acoustic control field.

The inverse design of non-parameterized ventilated acoustic resonators uses a different latent factorization. Each resonator is represented by a 128×64 binary cross-section image and a 50-point STL response vector sampled from 1 to 1961 Hz at 40 Hz steps. The AR-VAE defines a stochastic geometry latent $z$4, a deterministic geometry–acoustics alignment latent $z$5, and a response latent $z$6, with the decoder taking $z$7 to reconstruct geometry (Cho et al., 2024). Inverse design proceeds by encoding a target response $z$8 to $z$9, sampling $P$0 values of $P$1, decoding candidate images, thresholding them, and selecting the geometry with minimum STL MSE after COMSOL evaluation. For six representative targets, reported per-target MSE values are, for example, 0.966 versus 16.477 versus 131.981 at 601 Hz and 0.205 versus 9.882 versus 36.583 at 801 Hz for AR-VAE, best-in-train, and APNN, respectively (Cho et al., 2024). Averaged across these targets, AR-VAE reduces the MSE by roughly 25× compared to APNN and by about 2.5× compared to the closest design present in the training set (Cho et al., 2024). This suggests a control-oriented interpretation of LAM: the latent space binds acoustic response and structural realizability while preserving one-to-many design diversity.

6. Cross-modal mappings: articulation and safety alignment

Some recent work uses LAM to connect acoustic latents to variables that are neither spatial nor propagative. In acoustic-to-articulatory inversion, LAM is the learned regression

$P$2

where $P$3 for the $P$4-VAE and is reduced to 64 for EnCodec and Wav2Vec through spherical interpolation-based dimensional adjustment (Cámara et al., 2024). Audio is segmented into 15 ms frames, converted to 128-bin mel-spectrograms, and processed either by a jointly trained two-head $P$5-VAE or by frozen foundation encoders followed by the same projector MLP. The total training loss augments the VAE objective with a weighted squared regression loss on current articulatory parameters and a Huber smoothness loss on the previous frame. Across static and dynamic Pink Trombone datasets, the paper reports normalized median error < 0.1 for all cases on PT inputs, average MOS ≈ 3 for all three encodings in dynamic experiments versus human recordings, and a speed difference of approximately three orders of magnitude relative to classical optimization-based inversion, with neural inference near real-time versus approximately 15 minutes for 1 s audio (Cámara et al., 2024). In this formulation, the latent space is valuable because it is simultaneously acoustically reconstructive and articulatorily predictive.

A much more controversial usage appears in security research on Large Audio LLMs. There, LAM is described as the mapping from an input waveform into Acoustic Latent Semantics, defined as “the underlying paralinguistic features intrinsic to the priors of audio generative models,” mined from Bark’s native “history prompt” latent (Wang et al., 18 May 2026). The paper argues that instruction-neutral audio carrying specific ALS can induce “directional inference path drift” away from the safety alignment subspace. On JBB, selected ASR-M results include 100.00 versus 50.98 for Qwen2.5-Omni, 96.30 versus 61.11 for Qwen2-Audio, and 100.00 versus 50.00 for Gemini-2.5-Pro for AIA versus text-only; on WildJailbreak, selected results include 95.77 versus 47.89 for Qwen2.5-Omni and 85.71 versus 37.14 for Gemini-3-Pro (Wang et al., 18 May 2026). This usage is conceptually distant from localization or inverse design, but it still conforms to the same broad template: acoustic signals are mapped into latent variables that exert structured downstream effects not reducible to lexical content alone.

7. Evaluation practices, limitations, and unresolved standardization

LAM papers do not share a common evaluation protocol, because they do not share a common object of prediction. In-the-wild scene analysis relies on cosine distance heatmaps, t-SNE visualizations, and qualitative cluster emergence under pseudo-labeled GPS cells (Montero-Ramírez et al., 2024). LOCA-based mapping reports MAE after Procrustes alignment to physical coordinates (Cohen et al., 2023). OPTMA-Net uses normalized MSE and relative error on SPL, with separate generalization and extrapolation splits (Iqbal et al., 2022). Task-agnostic room estimation uses MAE, PCC, and BIAS for $P$6 and $P$7 across octave bands (Götz et al., 2024). Acousto-thermal mapping reports PSNR, RMSE, and SSIM on temperature fields (Cengiz et al., 2024). VAR inverse design uses MSE between simulated and target STL curves (Cho et al., 2024). Self-supervised DoAE uses Localization Error and Localization Recall (Roman et al., 8 Jul 2025). Reciprocal Latent Fields combine objective MAE on acoustic parameter fields with a MUSHRA-like listening test (Seuté et al., 6 Feb 2026). This diversity makes direct cross-paper comparison inappropriate unless the underlying latent semantics are first matched.

The limitations are likewise heterogeneous. In the WE-LIVE study, GPS is noisy and sparse, labels are pseudo-labels, event detection can be biased, and cross-user generalization is not evaluated (Montero-Ramírez et al., 2024). LOCA assumes a stationary environment, fixed sources, dense burst sampling, and feature smoothness; single-source training can deform embeddings at region-of-interest edges (Cohen et al., 2023). OPTMA-Net relies on a low-fidelity monopole approximation, free-field propagation, and identifiable source parameterization (Iqbal et al., 2022). The AIR-latent approach depends on the AIR-VAE training distribution and is trained on clean reverberant speech without additive noise (Götz et al., 2024). The AHL framework assumes linear acoustics and small-signal absorption and uses phase-only holograms with quantization imposed by printer resolution (Cengiz et al., 2024). AR-VAE omits a physics-informed training loss, models lossless air, and can decode implausible geometries for extreme latent samples (Cho et al., 2024). In DoAE, Localization Recall can be sensitive to domain shift when a supervised head is attached to UpLAM, although validation enrichment improves performance (Roman et al., 8 Jul 2025). RLF is trained per static map and does not handle dynamic geometry (Seuté et al., 6 Feb 2026).

The resulting picture is that LAM is not a settled subfield with a canonical benchmark, but a recurring design principle in acoustic machine learning and computational acoustics. A plausible implication is that the term will remain useful only if accompanied by an explicit statement of what the latent variable encodes and what structure constrains it: scene context, geometry, source physics, room identity, articulatory state, reciprocal propagation, or controllable holographic field. Without that qualification, “Latent Acoustic Mapping” is too broad to be self-explanatory. With it, the term names a coherent research strategy: encode acoustically meaningful variables in a latent domain whose geometry is easier to regularize, decode, and reason about than the original signal space.