We propose a machine-learning method that predicts trained-listener panel ratings of vehicle audio systems from in-situ microphone-array measurements. Using a dataset collected over more than 10 years from 1177 vehicle audio systems, the model is trained to predict listener-score distributions, not only mean scores, for subjective sound-quality attributes on a 1–10 scale. In this paper we focus on the timbral/spectral attribute and evaluate performance on a held-out set of 237 systems. Results show that prediction accuracy improves consistently with dataset size and that distribution-aware modelling captures not only expected score level but also listener disagreement and uncertainty. The study highlights both the feasibility of virtual listener panels and the importance of large-scale, reliable benchmarking data.
