Predicting Musical Meaning in Audio Branding Scenarios

This paper describes the concept of applying automatic music recommendation to the audio branding domain. We describe our approach of developing a prediction model for the perceived expressive content of music which is based on a large-scale listening experiment. We present an orthogonal 4-factor model for measuring musical expression as outcome variable, whereas audio- and music features as well as lyric-based features are introduced as prediction variables in the model. Furthermore, we describe Random Forest Regression as a concept for feature selection required to develop a Multi-Level Regression Model, which is taking individual listener parameters into account. Finally, we present first results from a preliminary stepwise regression model for perceived musical expression.