Toward a computational model of perceived spaciousness in recorded music

A computational model of perceived spaciousness in recorded musical signals, inspired by research in music information retrieval (MIR), is presented and evaluated experimentally. First three dimensions of spaciousness are selected for computational modeling: width of source ensemble, extent of reverberation, and extent of immersion. Next human subject responses to stereophonic music along these dimensions were collected. Then audio features from this data set of music were extracted and finally three exemplar-based machine learning models for estimating the three dimensions of spaciousness were trained and tested. The worst predictor was found to perform at least 32% better than a baseline predictor. These results are important to the music and audio engineering communities, as computational models for spaciousness in music may present new, perceptually meaningful methods of analysis and processing for the spatial characteristics of recorded music.