Contribution
Railway Psychoacoustic Annoyance: Custom Time-varying Psychoacoustic Metric Design with CRNNs
* Presenting author
Abstract:
We present a custom, time-varying, and linearly scaled psychoacoustic metric that captures the unique spectro-temporal characteristics of railway noise and correlates better with human perception of annoyance than the traditional LAeq rating. Our research is part of the EAV-Infra project (www.eav-infra.de), which investigates the potentials of Building Information Modelling, auralization, and visualization for infrastructure planning.Binaural recordings of 335 train passings were performed, and 50 recordings were presented in a listening test to 22 participants who rated time-varying annoyance on a scale from 0 to 100. A Convolutional Recurrent Neural Network (CRNN) was trained with mel-log spectra of the passings to predict the median time-varying annoyance ratings.Key findings reveal an RMSE of 5.93 between listening test ratings and predictions, well within the inter-personal interquartile range. Pre-training the network to predict the established Widmann Psychoacoustic Annoyance (PA) using a larger dataset with all available audio recordings achieved a promising RMSE of 1.24.The Railway Psychoacoustic Annoyance (RPA) indicator is proposed as a valuable addition for optimizing noise control measures in infrastructure, especially when competing options yield similar level reductions. Integration into planning software or audiovisual simulation applications is suggested to visually represent and assess annoyance in a more accessible way.