AcousticsML: ML in Acoustic Systems

• AcousticsML is a multidisciplinary field that merges acoustics, signal processing, and machine learning for tasks like classification, regression, and generative design.
• It provides reproducible, open-source Jupyter-based workflows and physics-informed models to address challenges in bioacoustics, spatial acoustics, and room acoustics.
• The domain balances data-driven flexibility with physical constraints, applying surrogate models and inverse design to optimize acoustic simulations and edge deployments.

AcousticsML denotes a body of work at the intersection of acoustics, signal processing, and machine learning, and in one explicit formulation it is presented as both a review of machine learning in acoustics and an open-source, Jupyter-notebook-based resource for reproducible workflows. In that formulation, the field spans classification, regression, and generation in acoustic data, with examples including acoustic data classification, generative modeling for spatial audio, and physics-informed neural networks (McCarthy et al., 6 Jul 2025). Across related literature, the same umbrella encompasses data-driven analysis of acoustic and musical signals, room and spatial acoustics, bioacoustics, ecoacoustics, distributed sensing, acoustic materials, and inverse design (Bianco et al., 2019).

1. Scope and intellectual framing

Machine learning in acoustics is repeatedly defined in contrast to conventional acoustics and signal processing. The 2019 review on the subject describes ML as data-driven, emphasizes that acoustic data provide scientific and engineering insight in fields ranging from biology and communications to ocean and Earth science, and argues that, given sufficient training data, ML can discover complex relationships between features and desired labels or actions, or between features themselves (Bianco et al., 2019). The later AcousticsML review extends that framing into a practical program: it provides Python-based demonstrations of classification, regression, generation, implicit neural representations, physics-informed neural networks, uncertainty quantification, and explainable AI, explicitly aimed at reproducible acoustic workflows (McCarthy et al., 6 Jul 2025).

A related conceptual precursor is the manifesto on machine listening intelligence, which defines an integrated framework for acoustic and musical signal modeling based on signal processing, deep learning, computational musicology, and representation theory (Cella, 2017). That perspective is important because it treats acoustic learning not as a single task family, but as a multilevel problem involving low-level, middle-level, and high-level representations. A plausible implication is that AcousticsML is best understood not as a single method class, but as a research program organized around representations, inference, and transformation.

Domain	Representative contribution	Paper
Bioacoustic frameworks	Open-source benchmark for cetacean call detection	(Bressler et al., 2023)
Ecological soundscapes	Coarse classification of biophony, geophony, anthropophony	(Gebhard et al., 20 May 2026)
Spatial acoustics	Physics-informed sound field estimation	(Koyama et al., 2024)
Room/acoustic transfer	One-shot matching of recordings across rooms or halls	(Verma et al., 2022)
Inverse design	Cloaks, coatings, and ultra-open silencers	(Tran et al., 2021, Weeratunge et al., 2022, Yang et al., 15 May 2026)
Edge acoustics	Tiny recurrent and TinyML deployment on low-power devices	(Rawat et al., 2021, Huang et al., 2024)

2. Representational and mathematical foundations

A recurrent formal starting point is the supervised-learning relation $y = f(x) + \epsilon$, where acoustic inference is cast as learning a mapping from measured inputs to labels, parameters, or reconstructed fields (McCarthy et al., 6 Jul 2025). The broader reviews then expand this into the standard divisions of supervised learning, unsupervised learning, semi-supervised learning, regression, classification, and generation (Bianco et al., 2019). In acoustics, the choice of representation is unusually consequential because raw waveforms, spectrograms, cepstral features, latent embeddings, and physically constrained fields are not interchangeable abstractions.

The machine listening intelligence perspective states four desirable properties for acoustic and musical representations: discriminability, stability, invariance to a transformation group $G$, and reconstruction (Cella, 2017). Those criteria recur implicitly across later AcousticsML systems. In practical notebook examples, the AcousticsML repository shows that feature choice can materially affect outcome: on AudioMNIST, FFT features with a random forest achieve accuracy 0.93, while MFCC features with a random forest achieve accuracy 0.99 (McCarthy et al., 6 Jul 2025). This is a compact demonstration that learned performance depends jointly on model family and acoustic representation.

Physics-informed formulations provide a second foundation. For sound field estimation in a source-free region, the field must satisfy the wave equation in time domain and the Helmholtz equation in frequency domain, with the paper writing the latter as

$$\left(\nabla_{\mathbf{r}}^2 + k^2\right)u(\mathbf{r},\omega)=0.$$

That review surveys plane-wave expansion, spherical wave function expansion, equivalent source distribution, kernel ridge regression with Helmholtz-consistent kernels, physics-constrained neural networks, and PINNs, and argues that purely data-driven interpolation can fit microphone samples yet still generate physically implausible sound fields (Koyama et al., 2024). This establishes a characteristic AcousticsML tension: flexibility versus physical admissibility.

3. Perception, timbre, and psychoacoustic modeling

A perception-centered branch of AcousticsML uses machine learning not merely to classify sounds, but to relate measurable descriptors to listening-based distinctions. In the piano timbre development study, a concert grand piano was recorded at two stages—right after manufacture and after one year of regular performance in a concert hall—and 176 recordings $(2 \times 88$ keys) were analyzed with a two-dimensional Self-Organizing Map using psychoacoustic timbre features from the apollon and COMSAR MIR frameworks. Spectral flux alone perfectly clustered the two stages; SPL, roughness, and fractal correlation dimension ordered keys by register rather than stage; and the combinations spectral flux + SPL and spectral flux + fractal correlation dimension preserved stage clustering while revealing sub-clusters. The reported interpretation is that stage 2 shows SPL homogenization and more systematic organization of initial transient chaoticity (Plath et al., 2021). The significance is methodological: unsupervised topology-preserving maps were used to expose both global stage separation and keyboard-dependent structure.

A closely related direction appears in automotive sound quality for EVs, where standardized psychoacoustic metrics are coupled with lightweight ML baselines. That work implements loudness (ISO 532-1/2), tonality (DIN 45681), roughness, fluctuation strength, sharpness, psychoacoustic annoyance, and LUFS reporting via ITU-R BS.1770-5, and uses the feature vector $\mathbf{x} = [N,\, S,\, R,\, F,\, T,\, PA]^{\top}$ on synthetic EV-like cases. The synthetic classes are engine boom, wind whistle, and road noise, and with fixed splits, seed = 123, and Random Forest $(n\_estimators = 100)$, the classifier achieves 0.93 accuracy (Goswami, 21 Sep 2025). Because the work is explicitly framed as an end-to-end example rather than a benchmark, its importance lies less in leaderboard performance than in the reproducible coupling of psychoacoustic descriptors with interpretable ML.

These studies jointly indicate that AcousticsML often operates on perceptually structured descriptors rather than raw SPL alone. A common misconception is that acoustic ML necessarily abandons psychoacoustics in favor of opaque embeddings; these examples show the opposite tendency in several subfields.

4. Monitoring, ecology, and distributed acoustic sensing

Bioacoustics and ecological monitoring are among the most operationally mature AcousticsML domains. Soundbay is an open-source Python framework for bioacoustics built on PyTorch with Hydra-managed configuration files, modular separation of model, optimizer/loss, and data pipeline, and benchmark support for cetacean call detection across OrcaData, ICML2013 Right Whale Redux, and a Mozambique / Deep Voice 2018 humpback whale dataset. In its reported benchmark, the Default model—ResNet50 with Adam—achieves AUC 0.915 on OrcaData, 0.984 on ICML2013, and 0.966 on the Deep Voice dataset, and the framework exports CSV outputs and annotation files compatible with Raven Lite/Pro (Bressler et al., 2023). The broader significance is infrastructural: reproducible cross-dataset comparison becomes part of the acoustics workflow rather than an afterthought.

On-device bioacoustics imposes a different constraint regime. TinyChirp studies bird-song recognition on low-power wireless acoustic sensors using a released dataset with 4737 target and 9395 non-target 3-second segments, downsampled to 16 kHz. It compares spectrogram-based and raw-audio TinyML architectures on an nRF52840 Development Kit with 1 MB Flash and 256 kB RAM. CNN-Mel achieves about 0.99 accuracy and 0.99 $F_2$ on the test set, while Transformer-Time achieves 0.93 accuracy and 0.91 $F_2$; however, once on-device preprocessing is included, spectrogram computation dominates cost, and the paper concludes that raw-audio models are preferable when end-to-end on-device cost matters. In the reported Corn Bunting deployment scenario, screening can save about 90% of SD-card space and extend storage-limited maintenance intervals from about 2 weeks to nearly an entire season of 18 weeks (Huang et al., 2024). This makes deployment economics a first-class modeling criterion.

CoarseSoundNet addresses ecological soundscapes at a coarser semantic level. It treats recordings as multi-label mixtures of biophony, geophony, and anthropophony, evaluates CNNs, transformers, and foundation-model-based encoders, and shows a domain gap between Edansa-2019 and BEsound. On Edansa, CNN10 reaches macro F1 .925; on BEsound, transformer/foundation-model-based systems such as Qwen2-Audio are stronger initially. The study further reports that adding an explicit silence class improves BEsound performance, and that post hoc proportional duration annotation plus class-specific thresholds raises macro F1 from .758 to .797, with count-based thresholding reaching .799 (Gebhard et al., 20 May 2026). The ecological implication is pragmatic: coarse soundscape filtering can become a preprocessing layer for downstream acoustic-index analysis.

Distributed acoustic sensing extends acoustic monitoring beyond conventional microphones. In the DAS benchmark, raw strain-rate data are converted into Cross-Spectral Density Matrix phase maps, augmented to 22,365 samples, and classified into vehicle noise, random noise, slope failure, and narrow-band noise. ResNet achieves accuracy 0.958 and F1-score 0.932, the best overall result, while SVM is the strongest classical baseline at accuracy 0.950 and F1-score 0.913 (Shi et al., 26 Mar 2025). This benchmark is notable because it treats fiber-optic sensing as an acoustics-recognition problem and makes explicit the trade-off between deep-model accuracy and deployment complexity.

Automated wildlife density estimation introduces a different statistical layer: integrating ML uncertainty into inference rather than using classifier outputs as fixed truth. In acoustic spatial capture-recapture for Hainan gibbons, ignoring false positives yields 17% positive bias, while three confidence-aware likelihood constructions give negligible bias and coverage probabilities close to the nominal 95% level (Wang et al., 2023). This is a technically important step because it inserts classifier uncertainty directly into ecological estimation.

5. Spatial acoustics, room inference, and acoustic transfer

Room and spatial acoustics have become a major AcousticsML test bed because they combine rich physics, expensive simulation, and sparse measurements. In early-stage acoustic design assessment, a parametric shoebox room was simulated in Pachyderm acoustic software across 2916 configurations, producing 72,900 data points after octave-band outputs were included. Seven fully connected DNN models were trained from geometric and material inputs to predict T30, EDT, C80, D50, and STI, achieving average error between 1% and 3% on model evaluation and between 2% and 12% on a separate unseen validation set of 48 room configurations; SHAP analysis identified wall absorption coefficient and room dimensions as among the most influential parameters (Abarghooie et al., 2021). The work shows how surrogate models can compress a 45-day simulation campaign into a fast decision-support tool.

A more generative room-related problem is one-shot acoustic matching. The proposed non-autoregressive Transformer-based system takes a source spectrogram and a single conditioning audio example from a target room or hall, extracts an acoustic signature with a CNN encoder, predicts a residual spectrogram, and reconstructs audio with Griffin–Lim. The architecture uses input/output spectrograms of size $257 \times 300$, embedding size 257, feed-forward dimension 512, 8 attention heads, and 3 layers, and it is trained on about 300,000 three-second patches from about 20 hours of audio. The reported evaluator, a Siamese CNN based on EfficientNet-B0 embeddings, has about 94% accuracy; human evaluation reports MOS 3.6 for predicted audio versus 4.1 for ground truth, with 76% correct judgment of closeness to target (Verma et al., 2022). This reframes room transfer as a conditional residual-learning problem rather than explicit impulse-response measurement.

The physics-informed sound field estimation literature occupies the most explicitly PDE-constrained end of this spectrum. It formulates interior sound field estimation as interpolation from sparse microphone samples in a source-free region, but argues that generic interpolation is structurally inadequate unless the governing physics are incorporated. The surveyed methods include linear expansions in physically valid basis functions, sparse dictionaries, Helmholtz-consistent kernels such as $$\kappa(\mathbf{r},\mathbf{r}') = j_0\!\left(k\|\mathbf{r}-\mathbf{r}'\|\right)$$, physics-constrained neural networks that predict expansion coefficients, and PINNs that add PDE residuals to data loss (Koyama et al., 2024). This literature is central to AcousticsML because it formalizes the claim that acoustics is not merely another regression domain.

6. Inverse design, metamaterials, and materials informatics

Inverse design is one of the clearest demonstrations that AcousticsML extends beyond recognition into synthesis of acoustic structures. In acoustic cloak design, the objective is to minimize total scattering cross section $G$0 for 2D configurations of rigid cylindrical scatterers. The study uses $G$1 binary images as inputs, predicts TSCS at 11 discrete wavenumbers, finds that CNNs outperform fully connected networks for the forward problem, and then combines VAE-based latent representations with Gaussian-process optimization for inverse design. The method is demonstrated for up to eight cylinders, with low-TSCS structures found in a few minutes for $G$2 (Tran et al., 2021). The conceptual move is to shift from repeated full-wave search to learned latent-space optimization.

Underwater acoustic coatings provide a second materials-informatics example. A COMSOL-based FEM of polyurethane coatings with cylindrical voids and steel backing was used to generate about 150,000 data points from 400 combinations of geometric parameters over 10 Hz to 10 kHz. Separate DNN surrogates for PU80, PU65, and PU90 take 10 geometric parameters plus frequency as input and predict absorption coefficient with Pearson correlation coefficient $G$3 and MAPE around 1.22% training, 1.27% validation, and 1.27% testing. Because DNN prediction for a frequency sweep takes about 0.04 s versus about 180 s for FEM, the reported acceleration is about $G$4, after which a genetic algorithm searches for broadband low-frequency attenuation. For PU80, the optimized design shows peaks of 0.96 at about 970 Hz and 0.99 at about 3830 Hz (Weeratunge et al., 2022). Here, ML is explicitly embedded inside optimization rather than appended as a post hoc analyzer.

A more recent physics-aware inverse-design formulation targets ultra-open acoustic silencers. That framework uses Green’s-function-based parameterization to decouple spectral parameters $G$5 from radial parameters, trains a two-stage forward predictor on 200,000 design-response pairs, and applies a population-based, hybrid-objective parallel inverse strategy that identifies hundreds of optimized candidates within seconds. It further reports hidden linear rules in the optimized radial parameters, including $G$6 and $G$7, interpreted as geometric proxies for optimal impedance matching. Experimentally, UAS-2 reaches about 80% ventilation, while the overall framework reports broadband bandwidth exceeding 830 Hz with thickness about $G$8–$G$9 (Yang et al., 15 May 2026). This suggests a shift from surrogate acceleration alone to ML-assisted discovery of design rules.

7. Software ecosystems, deployment regimes, and recurring issues

AcousticsML has also developed as an infrastructure and reproducibility agenda. The AcousticsML repository is organized into six chapters—signal processing introduction, feature extraction and selection, unsupervised ML, supervised ML, deep learning applications, and explainable AI—and is explicitly designed as a tutorial-style resource in Python and Jupyter notebooks (McCarthy et al., 6 Jul 2025). Soundbay performs a similar infrastructural role for bioacoustics by standardizing experiment configuration, preprocessing, augmentation, training, evaluation, sweep-based optimization, and benchmark comparison (Bressler et al., 2023). In both cases, reproducibility is treated as part of the scientific contribution.

Edge deployment imposes a distinct design discipline. For autonomous vehicles, the FastGRNN-based acoustic AI module uses 13-dimensional MFCCs, spectral gating for noise reduction, and a 26-cell recurrent architecture to classify six classes—Car Horn, Children Playing, Dog Bark, Drilling, Engine Idling, and Siren. FASTGRNN reaches 87.89% accuracy with 1,230 parameters and a size of 4.8 KB, and is framed as deployable on Arduino Uno, ARM microcontrollers, and Raspberry Pi (Rawat et al., 2021). The central trade-off is explicit: slightly lower accuracy than larger DNN/GRU/LSTM models in exchange for dramatic savings in size and latency.

A common misconception is that AcousticsML is equivalent to deep end-to-end classification. The literature is more heterogeneous. It includes Self-Organizing Maps for timbre development (Plath et al., 2021), Random Forest regression for acoustic emission where limited data and feature interpretability motivated a simpler model (Berta et al., 2024), kernel and basis-expansion methods for sound field estimation (Koyama et al., 2024), and lightweight logistic regression, random forest, and SVM baselines for psychoacoustic EV sound quality and DAS recognition (Goswami, 21 Sep 2025, Shi et al., 26 Mar 2025). Deep learning often performs best in benchmark settings, but classical models remain relevant when interpretability, data volume, or deployment constraints dominate.

Recurring limitations are likewise consistent across domains. The major reviews emphasize that ML in acoustics is data-hungry, less interpretable than classical approaches, and constrained by the no free lunch theorem (Bianco et al., 2019). Domain mismatch appears in ecological soundscapes, where anthropophony and geophony degrade sharply across recording domains and annotation mismatch remains nontrivial (Gebhard et al., 20 May 2026). DAS benchmarks identify real-time efficiency, limited labeled data, environmental and seasonal variation, and class imbalance as deployment bottlenecks (Shi et al., 26 Mar 2025). Physics-informed work adds a more fundamental warning: sparse measurements and unconstrained interpolation can yield nonphysical estimates unless the governing structure of the acoustic field is encoded in the model (Koyama et al., 2024). The resulting picture is not of a settled toolbox, but of a technically diverse field in which representation choice, physical prior, benchmark design, and deployment context remain inseparable.