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

Abstract:

In the automotive audio sector, computational engine sound modeling is a cornerstone, especially for virtual prototyping and active sound design. However, the rise of data-driven engine sound synthesis techniques has created a demand for extensive libraries of high-quality, standardized audio recordings. These resources must feature precise, time-aligned annotations regarding operating states, yet acquiring such data is notoriously challenging due to the prohibitive costs, the need for specialized measurement hardware, and the presence of unavoidable noise contamination.

To address these limitations, we introduce an analysis-driven framework designed to produce engine audio paired with sample-accurate control annotations. This approach utilizes pitch-adaptive spectral analysis to isolate harmonic structures from authentic recordings, which subsequently guide an extended parametric harmonic-plus-noise synthesizer. By applying diverse control trajectories and parametric variations, the framework expands a source audio clip of just 5 to 10 minutes per engine by a factor of 15 to 30. This process yields the Procedural Engine Sounds Dataset, comprising 19.0 hours of audio across 5,935 files. This collection features engine signals annotated with exact RPM and torque values, covering a broad spectrum of operating conditions, signal complexities, and harmonic profiles.

Validation against actual recordings demonstrates that the synthesized data retains the characteristic harmonic structures of real engines. Furthermore, a baseline differentiable synthesis network trained on this dataset confirms its efficacy for data-driven engine sound modeling. To foster advancements in engine timbre analysis, control parameter estimation, and neural generative synthesis, the dataset has been made publicly available.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC