Analysis and prediction of the audio feature space when mixing raw recordings into individual stems

Abstract

Processing individual stems from raw recordings is one of the first steps of multitrack audio mixing. In this work, we explore which set of low-level audio features are sufficient to design a prediction model for this transformation. We extract a large set of audio features from bass, guitar, vocal and keys raw recordings and stems. We show that a procedure based on random forests classifiers can lead us to reduce significantly the number of features and we use the selected audio features to train various multi-output regression models. Thus, we investigate stem processing as a content-based transformation, where the inherent content of raw recordings leads us to predict the change of feature values that occurred within the transformation.