Automatic Mixing

Linear Dynamical Systems

This work performs a supervised learning task to learn mixing weights (fader values) for a set of multi-track data. Ground truth mixing coefficients are estimated for 48 songs selected from the RockBand video game. These weights are used to train linear dynamical systems to predict mixing coefficients for an unknown set of tracks from acoustic features extracted from the audio.

Instrument Informed Mixing

Although digital music production technology has become more accessible over the years, the tools are complex and often difficult to navigate, resulting in a large learning curve for new users. Here, we approach the task of automated multi-track mixing from the perspective of applying common practices based on the instrument types present in a mixture. We apply basic principles to each track automatically, varying the parameters of gain, stereo panning, and coarse equalization. Assuming all instruments are known, a small listening evaluation is completed on the mixed tracks to validate the assumptions of the mix- ing model. This work represents an exploratory analysis into the efficacy of a hierarchical approach to multi-track mixing using instrument class as a guide to processing techniques.

Multi-Track Dataset

This publicly available dataset consists of 135 multi-track sessions with labels of the instrument present on each track.