Files

Abstract

Target-based assisted orchestration can be thought of as the process of searching for optimal combinations of sounds to match a target sound, given a database of samples, a similarity metric, and a set of constraints. A typical solution to this problem is a proposed orchestral score where candidates are ranked by similarity in some feature space between the target sound and the mixture of audio samples in the database corresponding to the notes in the score; in the orchestral setting, valid scores may contain dozens of instruments sounding simultaneously. Generally, target-based assisted orchestration systems consist of a combinatorial optimization algorithm and a constraint solver that are jointly optimized to find valid solutions. A key step in the optimization involves generating a large number of combinations of sounds from the database and then comparing the features of each mixture of sounds with the target sound. Because of the high computational cost required to synthesize a new audio file and then compute features for every combination of sounds, in practice, existing systems instead estimate the features of each new mixture using precomputed features of the individual source files making up the combination. Currently, state-of-the-art systems use a simple linear combination to make these predictions, even if the features in use are not themselves linear. In this work, we explore neural models for estimating the features of a mixture of sounds from the features of the component sounds, finding that standard features can be estimated with accuracy significantly better than that of the methods currently used in assisted orchestration systems. We present quantitative comparisons and discuss the implications of our findings for target-based orchestration problems.

Details

Actions

Preview