Automatic Music Mixing Using a Generative Model of Effect Embeddings

Eloi Moliner 1, Marco A. Martínez-Ramírez 2, Junghyun Koo 2, Wei-Hsiang Liao 2, Kin Wai Cheuk 2, Joan Serrà 2, Vesa Välimäki 1, Yuki Mitsufuji 2,3

1 Acoustics Lab, Department of Information and Communications Engineering, Aalto University, Finland
2 Sony AI
3 Sony Group Corporation

Listening Test Examples


Example Equal Loudness E2E-Flow FxNorm-AutoMix L MEGAMI I-L (proposed) Human Dry Stems
Vocals Bass Drums Other
2 - Rock
3 - Dance
4 - Disco
5 - Country
6 - Grunge
7 - BritPop

Mixing Variations


MEGAMI is a probabilistic model, which allows us to sample different mixing variations. Some examples are shown below.

Example
2 - Rock
3 - Dance
4 - Disco
5 - Country
6 - Grunge
7 - BritPop