Analysis of Acoustic Features for Automated Multi-Track Mixing

The capability of the average person to generate digital music content has rapidly expanded over the past several decades. While the mechanics of creating a multi-track recording are relatively straightforward, using the available tools to create professional quality work requires substantial training and experience. We address one of the most fundamental processes to creating a finished product, namely determining the relative gain levels of each track to produce a final, mixed song. By modeling the time-varying mixing coefficients with a linear dynamical system, we train models that predict a weight vector for a given instrument using features extracted from the audio content of all of the tracks.