Predicting the Perceived Spaciousness of Stereophonic Music Recordings

In a stereophonic music production, music producers seek to impart impressions of one or more virtual spaces upon the recording with two channels of audio. Our goal is to map spaciousness in stereophonic music to objective signal attributes. This is accomplished by building predictive functions by exemplar-based learning. First, spaciousness of recorded stereophonic music is parameterized by three discrete dimensions of perception—the width of the source ensemble, the extent of reverberation, and the extent of immersion. A data set of 50 song excerpts is collected and annotated by humans for each dimension of spaciousness. A verbose feature set is generated on the music recordings and correlation-based feature selection is used to reduce the feature spaces. Exemplar-based support vector regression maps the feature sets to perceived spaciousness. We show that the predictive algorithms perform well on all dimensions and that perceived spaciousness can be successfully mapped to objective attributes of the audio signal.