AI4TEN: Synthetic-to-Real Transfer for Acoustic Vehicle Classification Using Physics-Based and AI-Generated Training Data

This repository contains the data, code, and trained models for the study "AI4TEN: Synthetic-to-Real Transfer for Acoustic Vehicle Classification Using Physics-Based and AI-Generated Training Data." DOI: https://doi.org/10.3390/app16147234. Weblink: https://www.mdpi.com/2076-3417/16/14/7234. The study compares two synthetic data generation approaches for training a CNN-based vehicle classifier that separates cars, trucks, and motorcycles. The first approach uses physics-based simulation with pyroadacoustics. The second uses AI text-to-audio generation with AudioLDM. A cross-dataset protocol trains on one dataset of 1,006 real samples supplemented with synthetic data and tests on two independent datasets totalling 1,019 samples recorded in different countries with different microphones. The evaluation covers eight training configurations across five random seeds. Combining real data with both synthetic sources reaches a balanced F1 of 0.39, a 55 percent improvement over the real-only baseline. A matched sample-size experiment shows that the two synthetic sources perform comparably as standalone training data. When combined with real recordings, physics-based augmentation keeps a modest advantage. A spectral analysis shows that AudioLDM produces sounds closer to real recordings, while combining both sources gives the largest gain. A congestion stress test shows a 43 percent F1 drop with two overlapping vehicles, alongside rising model confidence. Repository structure:- configs: experiment configuration files- data: real and synthetic audio datasets for training and evaluation- models: trained classifier weights- results: evaluation outputs, metrics, and figures- scalers: fitted preprocessing scalers- scripts: code to reproduce preprocessing, training, evaluation, and analysis See README for setup, dependencies, and instructions to reproduce the results.