PaperlayerView exact revision on arXiv

Source-linked AI summary

Supervised Source Separation of Electric Vehicle Interior Noise Using a Hybrid Transformer Architecture

Philipp Diel, Zhuo Cao, Arne Hildenbrand, Marco Günther, Stefan Pischinger, Hanno Scharr, Steffen Marburg

arXiv:2610.32310v1cs.SDcs.LGeess.AS

TL;DR

EV interior-noise contribution analysis requires separating wind, tire, powertrain, and auxiliary sounds, while obtaining measured training stems is costly. The paper trains a conditioned hybrid separator on labeled physics-based synthetic data and evaluates transfer to measured vehicles. Synthetic training learns audible or spectrally visible contributions, but measured-noise transfer remains incomplete without fine-tuning.

  • Problem

    Separating EV interior noise into source contributions is useful for acoustic development, but obtaining measured training data with known contributions requires costly transfer-path-analysis campaigns.

    Source: §1 · pp. 1–2

  • Method

    The paper trains conditioned and unconditioned hybrid time-spectrogram transformers on physics-based synthetic binaural mixtures with four labeled source contributions.

    Source: §6 · p. 16

  • Results

    Synthetic training effectively separates contributions that are audible or visible in spectrograms, but domain-randomized synthetic data alone is insufficient for measured vehicle noise.

    Source: §6 · p. 16

  • Takeaways & Limitations

    Synthetic training can provide a pretrained model for fine-tuning on measured data, which may reduce the amount of measurement data required.

    Source: §6 · p. 16

    Transfer to measured vehicle noise is partial: domain randomization does not close the simulation-to-reality gap, and errors include misassigned orders and hallucinated powertrain content.

    Source: §5.1 · p. 10

Abstract

from arXiv · show

The interior sound of an electric vehicle is a superposition of several distinct noise contributions - wind noise, tire noise, powertrain (electric drive unit and inverter) noise, and auxiliary-drive noise - whose relative balance governs both the overall loudness and the perceived sound quality in the vehicle interior. Decomposing a recorded vehicle interior signal into these source contributions is a prerequisite for targeted acoustic engineering, but requires effortful measurement campaigns as well as expertise in the field of transfer path analysis. This work formulates the decomposition as a supervised, binaural source-separation problem with four target source contributions and adapts a hybrid time-spectrogram transformer architecture, originally developed for music demixing, to the domain of sound source separation for technical systems. To disambiguate source contributions, conditioning is introduced in the form of operating-state information, including electric drive speed, electric drive load, vehicle speed, and auxiliary-drive speed. Since isolated real-world reference stems are costly to obtain, the model is trained on a physics-based parametric data generator that synthesizes labeled component sounds from randomized virtual-vehicle hardware configurations and driving scenarios, so that the ground-truth contribution of each source is known by construction. Validation on both synthetic and measured vehicle interior sounds probes the achieved separation performance and underscores the feasibility of applying the proposed method for acoustic contribution analysis in technical systems.

1 Introduction

EV interior sound is shaped by the balance of wind, tire, powertrain, and auxiliary noise, making source attribution useful for acoustic development. The paper brings supervised audio separation and vehicle-noise contribution analysis together to test whether learned separation can reduce reliance on costly measured training data.

1 Introduction

  • Because EVs lack combustion-engine masking noise, wind, tire, powertrain, and auxiliary contributions govern interior sound and its perceived balance.Separating these contributions can support source-specific annoyance attribution, masking analysis, and assessment of design changes.Source: §1 · p. 1

1.1 Interior Noise of Electric Vehicles

  • Tire and wind noise are broadband sources, while powertrain and auxiliary components contribute tonal, order-harmonic noise; inverter noise is also relevant at high frequencies.Wind becomes prominent at high vehicle speeds, whereas auxiliary tonal noise can be especially audible at low speeds.Source: §1.1 · p. 2

1.2 Interior Noise Contribution Analysis

  • Established TPA distinguishes source strength from transfer paths, while OTPA estimates path contributions from operational measurements but can suffer from transmissibility, missing-path, and correlated-indicator errors.Classical and component-based TPA require instrumented transfer paths and additional standstill measurements.Source: §1.2 · p. 2

1.3 Data-based Audio Source Separation

  • Audio source separation estimates constituent signals from a mixture; hybrid time-spectrogram models combine waveform and spectrogram processing, while band-split models target source-specific frequency ranges.HT Demucs uses parallel waveform and spectrogram branches with a cross-domain transformer, and ranked highest among open-source models in the cited music-demixing comparison.Source: §1.3 · pp. 3–4

1.4 Conditioning mechanisms for separation

  • The paper uses synchronized operating-state side information to constrain separation, a form of conditioning distinct from specifying a target source.Related examples include time-aligned scores and drone motor speed, while the paper motivates a domain-specific supervised separator for technical noise.Source: §1.4 · p. 4, §1.5 · p. 4

1.5 Design Choices and Contributions

  • The paper formulates EV interior-noise decomposition as supervised source separation and contributes a physics-based generator, conditioned model variants, and evaluation on production-vehicle measurements.Its stated contributions also include separation results, domain-specific evaluation practice, and discussion of operating-state conditioning for technical systems.Source: §1 · p. 1, §1.5 · p. 5

2 Methodology

The methodology synthesizes labeled binaural source contributions across randomized virtual vehicles and driving scenarios, then trains a conditioned or unconditioned hybrid separator. It uses source-weighted waveform and spectrogram losses, with separation quality assessed by SDR alongside spectral inspection.

2.1 Dataset

  • The generator samples virtual-vehicle hardware and binaural transfer functions, creates 10-second driving scenarios, then synthesizes wind, tire, powertrain, and auxiliary contributions.Broadband noise is split into tire and wind components using a modeled transfer function; powertrain and auxiliary sounds are synthesized as tonal sources.Source: §2.1 · p. 5

2.2 Problem formulation

  • The task maps a binaural mixture to simultaneous binaural estimates of four sources: wind, tire, powertrain, and auxiliary-drive noise.Sources are generated from randomized hardware and operating scenarios, providing known contributions for supervised training.Source: §2.1 · p. 5, §2.2 · p. 6

2.3 Network architecture and conditioning

  • A dual U-Net processes waveform and spectrogram representations in parallel, with a cross-domain transformer at the bottleneck and per-source decoders.Time-synchronized drive speed, load, vehicle speed, and auxiliary-drive speed provide physically grounded cues about source characteristics.Source: §2.3 · p. 6

2.4 Training objective

  • Training combines source-weighted time-domain L1 reconstruction with an auxiliary spectrogram loss, and optional normalization balances sources with strongly varying amplitudes.Normalization divides each source's waveform loss by its target's mean absolute amplitude to reduce the tendency to prioritize loud components.Source: §2.4 · p. 7

2.5 Evaluation metrics

  • Separation quality is measured with time-domain SDR, but the authors caution that it can favor estimates with deficient spectral content and complement it with spectrogram inspection.They also propose third-octave-band-level error as an acoustic-engineering metric for broadband contributions.Source: §2.5 · p. 8

3 Experimental Setup

The experiments compare model variants on single-vehicle data and test generalization across virtual vehicles, including transfer to measured vehicle-interior noise. A shared conditional HT Demucs backbone is evaluated with changes to conditioning, loss design, and spectrogram representation.

3 Experimental Setup

  • Two datasets support the experiments: varied driving conditions for one virtual vehicle and 300 virtual vehicles with varied hardware, the latter also used to test separation on measured vehicle noise.The single-vehicle dataset holds hardware parameters constant to compare model variants and conditioning; the multi-vehicle dataset tests generalization across vehicle configurations.Source: §3 · p. 8
  • Training samples are six-second segments from ten-second binaural signals, resampled from 44.1 kHz to 16 kHz, with accelerated scenarios capped at 100 km/h.The single-vehicle dataset has 9237 training and 93 validation samples; the multi-vehicle dataset has 6093 training and 2127 validation samples.Source: §3 · p. 8
  • All models use a lightweight conditional HT Demucs backbone with waveform and spectrogram branches and a cross-transformer bottleneck; conditioning, when enabled, is injected through FiLM.The backbone has three encoder-decoder stages in each branch, while Demucs augmentations and audio normalization are disabled.Source: §3 · p. 8
  • The model predicts magnitude-domain sources and applies one-step Wiener refinement in training and inference, with the residual branch disabled so mixture energy is assigned to modeled sources.The STFT window is 4096 and hop length 1024 unless otherwise stated.Source: §3 · p. 8
  • Waveform-loss weights are [1, 1, 20, 20] for tire, wind, powertrain/inverter, and auxiliary-drive sources, emphasizing the more heavily masked contributions.Spectrogram-loss variants use spectral-convergence-only STFT loss.Source: §3 · p. 8
  • Eight single-vehicle model variants isolate conditioning, waveform-loss normalization, and auxiliary spectrogram loss while holding other architecture and optimization settings fixed.Table 1 summarizes the variant factors; optimization uses Adam with a learning rate of 5 × 10−4 and trains for up to 120 epochs.Source: §3 · pp. 8–9, Table 1 · p. 9

4 Results

The model separates measured vehicle-interior noise, but results remain unsatisfactory, and performance varies with source masking and vehicle hardware. The experiments also identify limitations in conditioned separation and in using SDR alone to assess quality.

Evaluation

  • High SDR for wind and tire noise does not by itself establish separation quality: some estimates with favorable SDR contain spectral artifacts, so evaluation should also use listening and spectrogram analysis.Spectrogram-loss variants score low on time-series SDR because their optimization also targets spectrogram reconstruction.Source: §4 · p. 9

Single-vehicle source estimates

  • In the single-vehicle example, tire noise is reconstructed well at low frequencies while higher-frequency noise remains, and wind noise is reproduced particularly well below 2 kHz.Powertrain orders are estimated more accurately when electric-drive load is higher; auxiliary-drive noise is less accurate than the other sources.Source: §4 · p. 9

Conditioning

  • Conditioned variants have lower separation quality than unconditioned variants despite favorable SDR values in some configurations, and exhibit recurring 250 Hz energy bands.A smaller STFT hop size most effectively mitigates the artifacts, but increases computational cost and training time.Source: §4 · p. 10

Multiple virtual vehicles

  • Across multiple virtual vehicles, tire and wind noise are approximated very well, but powertrain artifacts increase and masked low-frequency orders are no longer inferred; auxiliary-drive estimates are harder amid greater training variation.The task is harder because source transfer functions, order spectra, and relative intensities differ across vehicles.Source: §4 · p. 10

Measured vehicle-interior noise

  • On measured vehicle noise, UC-BASE assigns inverter switching noise and most audible orders to the powertrain, but order-harmonic noise also appears erroneously in wind and tire estimates and source content is hallucinated.Switching frequencies are recognizable at approximately 7.5 kHz for Vehicle A and 5 kHz for Vehicle B; overall results remain unsatisfactory.Source: §4 · p. 10

5 Discussion

The discussion finds source separation feasible on a single synthetic vehicle, but more fragile across hardware and in transfer to measured sounds. It argues for spectral-aware training and evaluation while positioning the model as a lower-effort complement, not replacement, to established contribution analysis.

5.1 Feasibility and transfer

  • Single-vehicle separation is learnable, but hardware variation degrades estimates gradually and most strongly affects the fully masked auxiliary drive.The evaluated scenarios exclude full-load accelerations, when powertrain excitation is strongest and its orders are least masked.Source: §5.1 · p. 10, Figure 1 · p. 11, Figure 4 · p. 14
  • Transfer to measured vehicle noise is partial: domain-randomized virtual hardware does not close the simulation-to-reality gap, motivating measured training data with known contributions.The observed errors are systematic and attributed to the training-data distribution rather than separator capacity.Source: §5.1 · pp. 10–11

5.2 Implications for the training objective, evaluation, and model configuration

  • Waveform loss supports reconstruction, while auxiliary spectrogram loss better targets spectrally clean contributions; logarithmic magnitude loss overcorrects and harms overall separation.The per-source loss emphasizes high-energy low-frequency tire content, while a logarithmic term shifts emphasis toward low-level background and underfits high-energy regions.Source: §5.2 · p. 11
  • Time-domain SDR is insufficient alone: spectrally cleaner estimates can score lower, and favorable SDR can coexist with audible periodic artifacts.The authors suggest third-octave level errors for broadband sources and tracked-order level errors for tonal sources as complementary metrics.Source: §5.2 · p. 11
  • Operating-state conditioning did not improve on unconditioned variants and produced 250-Hz bands; magnitude prediction with Wiener filtering and fine frequency resolution were beneficial configuration choices.The conditioning-vector representation remains unexplored, and the authors note that conditioning may be redundant for a single vehicle.Source: Figure 3 · p. 13, §5.2 · p. 15

5.3 Relation to established methods and applicability

  • Once trained, the model needs a binaural receiver-position measurement and can substantially reduce measurement effort, but cannot replace TPA’s accuracy or path-ranking and source-strength capabilities.The authors also propose combining learned estimates for costly-to-instrument contributions, such as tire noise, with OTPA for remaining paths.Source: §5.3 · p. 15
  • Compared with image-based methods, the approach can recover masked order content without hardware generalization, learns broadband tire–wind decomposition from data, and avoids manual feature engineering.The cited image-based decomposition uses a frequency-dependent ratio that is identical across test vehicles and operating points.Source: §5.3 · p. 15
  • Applicability depends on training data with known source contributions, which may come from isolated receiver recordings, partial test rigs, or TPA-recovered operational measurements.Examples include separately measured wind-tunnel, tire-drum, and dynamometer noise, as well as cabin thermal-management components and production-hall machines.Source: §5.3 · pp. 15–16

6 Conclusion

The work formulates four-source EV interior-noise decomposition as supervised separation and finds that data-driven models can isolate sufficiently audible or spectrally visible contributions. Synthetic training alone is insufficient for measured noise, while the evaluation also requires more than time-domain SDR.

6 Conclusion

  • Time-domain SDR alone is insufficient: favorable scores can accompany spectrally deficient estimates, so spectrogram and auditory assessment should be complemented by acoustic-engineering metrics.The error of third-octave band levels is suggested for broadband contributions.Source: §6 · p. 16
  • The model effectively isolates technical-noise contributions when they are sufficiently audible or visible in the spectrogram.The decomposition targets wind, tire, powertrain, and auxiliary-drive contributions using conditioned and unconditioned hybrid time-spectrogram transformer variants trained on physics-based synthetic data.Source: §6 · p. 16
  • For measured vehicle interior noise, domain-randomized synthetic data is insufficient alone, but may pre-train a model for measured-data fine-tuning and reduce required measurement data.Source: §6 · p. 16
  • Operating-state conditioning did not yield the expected benefit; variable selection and representation, and how conditioning enters the model, remain primary research directions.Source: §6 · p. 16
  • Once trained, the method requires only a receiver-position measurement and complements transfer path analysis as a low-effort alternative where labeled source contributions are available.Source: §6 · p. 16

7 Appendix

The appendix tables cover randomized properties and distributions, and how vehicle operating parameters influence synthesized sound sources.

7 Appendix

  • Table 2 covers randomly sampled properties, their value ranges, and underlying distributions.Source: Table 2 · p. 22
  • Together, the tables address sampled-property settings and operating-parameter effects in the sound-generation setup.Source: Table 2 · p. 22, Table 3 · p. 23
  • Table 3 covers the influence of vehicle operation parameters on synthesized sound sources.Source: Table 3 · p. 23
Loading 2610.32310v1…