101 mixes : a statistical analysis of mix-variation in a dataset of multi-track music mixes

The act of mix-engineering is a complex combination of creative and technical processes; analysis is often performed by studying the techniques of a few expert practitioners, qualitatively. We propose to study the actions of a large group of mix-engineers of varying experience, introducing quantitative methodology to investigate mix-variation and the perception of quality. This paper describes the analysis of a dataset containing 101 alternate mixes generated by human mixers as part of an on-line mix competition. A varied selection of audio signal features is obtained from each mix and subsequent principal component analysis reveals four prominent dimensions of variation - `dynamics', `treble', `width' and `bass'. An ordinal logistic regression model suggests that the ranking of each mix in the competition was significantly influenced by these four dimensions. The implications for the design of intelligent music production systems are discussed