Pulse-Train-Resonator Model

Physics-Informed Neural Engine Sound Synthesis

The Pulse-Train-Resonator (PTR) represents an advanced deep learning architecture designed for synthetic engine sound generation. This model innovatively simulates combustion pressure pulses and exhaust propagation utilizing differentiable synthesis components. Unlike traditional methods that primarily focus on spectral characteristics, the PTR directly models the fundamental pulse structure of engine noises, providing a more accurate representation of engine sound dynamics.

Overview

Understanding that engine sounds derive from a sequence of combustion pressure pulses rather than continuous harmonic oscillations, the PTR method utilizes parameterized pressure waves aligned with engine firing patterns. These waves are propagated through differentiable Karplus-Strong resonators, effectively simulating exhaust acoustics. Trained on the Procedural Engine Sounds Dataset, PTR showcases its ability to generalize across various engine configurations, outperforming traditional harmonic-plus-noise baselines while utilizing identical network architectures.

Key Features

• Physics-informed pulse synthesis: Integrates thermodynamic pitch modulation alongside valve-dynamics environments for authentic sound reproduction.
• Differentiable Karplus-Strong resonators: Provides a robust framework for accurate modeling of exhaust system acoustics.
• Firing-order sequencing: Each cylinder is parameterized for precise sound synthesis.
• Broad generalization capabilities: Maintains effective performance across different engine types despite the constraints of architectural design.

Repository Structure

This repository is organized as follows:

ptr-model/
├── checkpoints/                            # Pre-trained model weights
│   └── 2025-08-31_models_and_weights.zip
├── configs/                                # Base configuration files
├── scripts/                                # Training and inference scripts
│   ├── train.py                            # Training pipeline with CLI
│   └── inference.py                        # Model inference with CLI
├── src/                                    # Source code
│   ├── audio/                              # Audio processing utilities
│   ├── data/                               # Data loading and processing
│   ├── models/                             # Model implementations
│   │   ├── hpn_model.py                    # Harmonic-Plus-Noise baseline
│   │   ├── hpn_synth.py                    # HPN synthesis modules
│   │   ├── ptr_model.py                    # Pulse-Train-Resonator model
│   │   ├── ptr_synth.py                    # PTR synthesis modules
│   │   └── model.py                        # Base model architecture
│   ├── training/                           # Training utilities
│   └── utils/                              # General utilities
├── pyproject.toml                          # Python dependencies and metadata
├── README.md                               # This file
└── LICENSE                                 # License file

Performance Evaluation

Evaluation results underline the effectiveness of the PTR model against a Harmonic-Plus-Noise (HPN) baseline. Across three engine configurations, the PTR model achieved:

• 5.7% improvement in total validation loss
• 21% improvement in audio reconstruction on previously unseen data content.

Dataset Information

The model utilizes the Procedural Engine Sounds Dataset, which includes:

• 19 hours of engine audio recorded under varying operational conditions.
• Time-aligned annotations for RPM, torque, throttle, and DFCO.
• Coverage across diverse engine configurations and acoustic scenarios, providing a comprehensive data resource for training.

Dataset access is available through:

• Zenodo: Procedural Engine Sounds Dataset
• Hugging Face: Hugging Face Datasets

Additional Information

For supplementary resources, including audio examples and detailed model outputs, please visit: Audio Examples.

This repository serves as a pioneering effort in integrating physics-informed models within deep learning frameworks for audio synthesis, enabling broader applications beyond automotive sound design.