We propose that applying source separation as a pre-processing step to computer-assisted orchestration will improve the orchestrations of multi-layered sounds. In order to test this, we generated 300 targets, where each target is a combination of 4 source sounds. Each target is first orchestrated without separation, creating the "full target orchestration." Then, each separation method is applied to the target, and four sub-targets are generated by each method. The subtargets are individually orchestrated with a subset of the orchestra, then recombined to make the "separated orchestration."
Below is a selection of targets and orchestrations from our experiments.
| Target | Full Target Orchestration | Separated Orchestration |
|---|---|---|