Analysis-Driven Procedural Generation of an Engine Sound Dataset with Embedded Control Annotations

The paper introduces an analysis-driven framework designed for procedural engine sound synthesis, which addresses limitations in existing audio resources by generating comprehensive, annotated datasets from minimal source material. This methodology is critical because it enables substantial data expansion while rigorously preserving the physical and acoustic characteristics unique to specific engines, thereby facilitating advanced research in automotive acoustics and machine learning.

Acoustic Signature Preservation

The core technical challenge addressed is ensuring that synthesized sounds maintain physical fidelity across varying operating conditions. The framework achieves this by verifying that fundamental acoustic behavior tracks operating state. Crucially, the synthesis process ensures that Engine-specific signatures are preserved, specifically noting the presence of a dominant 4th order at V8 firing frequency, [and] 1.5th order during engine-braking. Furthermore, the system guarantees that the magnitude evolution correspond[s] across the RPMtorque operating space, limiting variations in higher orders (>8) to only those reflecting parametric modifications that extend timbral diversity beyond source material.

Procedural Synthesis and Dataset Expansion

The framework’s primary output is a highly valuable, annotated corpus. The authors report presenting an analysis-driven pipeline that generates annotated datasets, achieving a massive expansion of 15–30× expansion while preserving engine-specific acoustic signatures. This resulting dataset is unique because it includes sample-accurate embedded annotations, which directly addresses documented deficiencies in current engine audio resources.

Key Research Applications

The structured nature and comprehensive annotation of the generated data enable several advanced research avenues, including:

• Inverse parameter estimation (predicting RPM/torque from audio for automatic annotation and NVH diagnostics).

• Data-driven synthesis development (learning parameter mappings without manual tuning).

• Systematic algorithm evaluation via controllable modifications.

The documented analysis-synthesis pipeline is designed to be transferable, allowing researchers to apply the framework to their own recordings for task-specific corpus generation.

Model Training and Validation

The efficacy of the data-driven approach is validated using neural synthesis models. Performance metrics demonstrate that Stable convergence at early stopping demonstrates suitability for data-driven methods. The model's robustness is further assessed by monitoring how error points shift across different content types; specifically, Esc points... shift earlier from A to C with increasing nondeterministic content, indicating the method’s ability to handle complex variations in source material.

The core scientific contribution lies in establishing a physics-informed, analysis-driven pipeline that separates fundamental acoustic behavior (the physics) from arbitrary spectral content (the noise/texture). This separation allows for generalization beyond the training data.

Here are three specific, high-impact improvements to existing AI systems:

Current Limitation Addressed: Purely data-driven models (like standard WaveNet or GANs) often fail to extrapolate realistically outside the bounds of the training dataset, especially when extrapolating physical states (e.g., high RPM cornering, engine braking).

Proposed Improvement: Implement a Variational Autoencoder (VAE) or Flow-based Model architecture where the latent space (z) is explicitly constrained and guided by known physical relationships.

• Mechanism: Instead of mapping z to raw audio samples, the decoder must be structured as a cascaded system:

1. Physical Feature Predictor: A small network predicts key time-domain parameters (e.g., fundamental frequency f 0, dominant harmonic orders H n, and mechanical excitation rates) based on the input control vector (RPM, Torque). This component enforces known physical laws (e.g., the relationship between crankshaft speed and harmonic peaks).

2. Synthesis Layer: A differentiable digital signal processing (DSP) module (as hinted by the paper's use of differentiable pulse-train synthesis) takes these predicted parameters and generates a structured, quasi-periodic backbone signal (Signal Core).

3. Stochastic Modulation Layer: A second generative network (e.g., a specialized WaveGAN trained only on residuals) learns the non-deterministic, high-frequency components (Noise Texture) that cannot be captured by the physics model (e.g., tire squeal, air intake whistle).

• Loss Function: The loss function must be composite: L = lambda 1 L Reconstruction + lambda 2 L PhysicsConstraint + lambda 3 L Perceptual. The L PhysicsConstraint term penalizes deviations from known relationships (e.g., if the predicted H 4 at V8 firing frequency deviates from the expected magnitude).

What the Improved AI System Can Do:

• Guaranteed Physical Plausibility: Generate engine sounds that are not only statistically realistic but are physically consistent across vast, unrecorded operating regimes (e.g., simulating a unique sound profile at 70% torque and 8500 RPM, even if that exact combination was never recorded).

• Controllable Variation: Allow direct manipulation of acoustic features (e.g., Increase the perceived mechanical resonance by X dB or Simulate the effect of a specific exhaust backpressure) by adjusting input controls, going beyond simple spectral blending.

• Mechanism: The network processes spectrograms or Mel-Cepstrum Coefficients derived from the audio stream. Instead of outputting a single class label, the final layers must output a continuous, time-series vector representing the state: [RPM(t), Torque(t), Gear Ratio(t)].

• Training Data: This requires using the annotated datasets generated by the proposed framework (Section 1) as ground truth. The network is trained to minimize the error between its predicted physical state vector and the known ground truth vector at every time step.

• Focus on Harmonic Tracking: Incorporate a dedicated sub-module that uses Fourier analysis to explicitly track specific, predictable harmonic orders (H 4, H 1.5) over time, using these tracked values as intermediate feature vectors for the regression task.

• Mechanism: This toolkit accepts structured JSON input defining required acoustic modifications (e.g., "event": "rain impact", "severity": 0.8, "frequency band": [100-500]). It then manages the synthesis pipeline:

1. It runs the Physics Core (Section 1) to generate the base signal for the required RPM/Torque.

2. It calculates the necessary spectral modification coefficients based on external inputs (e.g., rain sound samples).

3. It uses advanced signal processing techniques (like frequency-domain multiplication or wave-shaping filters) to blend these components seamlessly, ensuring that the added noise component does not introduce spurious harmonics that violate the core physics model's constraints.

• Key Feature: The system must include a Cross-Domain Style Transfer Module trained to apply acoustic textures from one source (e.g., passing train sound) onto another (e.g., engine sound), maintaining the fundamental harmonic structure while altering the timbre entirely—a controlled form of spectral metamorphosis.

Computational engine sound modeling is central to the automotive audio industry, particularly for active sound design applications and virtual prototyping. Emerging data-driven engine sound synthesis methods require large volumes of standardized, clean audio recordings with precisely time-aligned operating-state annotations: data that is difficult to obtain due to high costs, specialized measurement equipment requirements, and inevitable noise contamination. We present an analysis-driven framework for generating engine audio with sample-accurate control annotations. The method extracts harmonic structures from real recordings through pitch-adaptive spectral analysis, which then drive an extended parametric harmonic-plus-noise synthesizer. With this framework, we augment 5-10 min of source audio per engine 15-30x via diverse control trajectories and parametric variation, producing the Procedural Engine Sounds Dataset (19.0 h, 5,935 files): a set of engine audio signals with sample-accurate RPM and torque annotations spanning a wide range of operating conditions, signal complexities, and harmonic profiles. Comparison against real recordings validates that the synthesized data preserves characteristic harmonic structures, and a baseline differentiable synthesis network trained on the dataset confirms its suitability for data-driven engine sound modeling. The dataset is released publicly to support research on engine timbre analysis, control parameter estimation, and neural generative synthesis.