Learning music production practice through evolutionary algorithms

The field of intelligent music production has been an active research topic for over a decade. The aim is to develop systems which are capable of performing common tasks in music production, such as level-balancing, equalisation, panning, dynamic range compression and application of artificial reverberation. Many systems developed are modelled as expert systems, where the music production task is solved by optimisation, and domain knowledge, obtained by examining industry “best-practice” methods, is used to determine the optimisation target [1]. Drawbacks to this method include the fallibility of domain knowledge and the assumption that there is a global optimum – a mix which all users would agree is best. Results suggest that many systems can perform basic technical tasks but struggle to compete with human-made mixes, due to a lack of creativity. We propose to use interactive evolutionary computation to solve this problem. These methods are well suited to aesthetic design problems, which are highly non-linear and non-deterministic. In the case of music mixing, the problem is highly subjective: research has shown that mix engineers typically prefer their own mix to those of their peers [2]. Consequently, intelligent music production tools would benefit from interactivity, to determine “personal” global optima in the solution space, instead of one “universal” global optimum. The space to be explored is a novel “mix-space” [3]. This space represents all the mixes that it is possible to create with a finite set of tools. Currently, basic level adjustment has been implemented, while mix-space representations of panning and equalisation are currently under development. The fitness function for optimisation is subjective, allowing mixes to be generated in accordance with any perceptual description, such as “warmth”, “punchiness” or “clarity”. Clustering techniques are used to increase the population size beyond that which a user could realistically rate, by extrapolating the fitness function to nearby individuals. When optimising the overall “quality” of the mix, we introduce findings from recent, large- scale studies of music mixes [4], which can be used to calculate the fitness of the population, alongside the subjective rating. Early results indicate that the system can produce a variety of mixes, suited to varying personal taste. We believe this approach can be used to further the study of intelligent music production, to deliver personalised object-oriented audio and increase the understanding how music is mixed