Music Production Behaviour Modelling
View/ Open
Metadata
Show full item recordAbstract
The new millennium has seen an explosion of computational approaches to the study of music production, due in part to the decreasing cost of computation and the increase of digital music production techniques. The rise of digital recording equipment, MIDI, digital audio workstations (DAWs), and software plugins for audio effects led to the digital capture of various processes in music production. This discretization of traditionally analogue methods allowed for the development of intelligent music production, which uses machine learning to numerically characterize and automate portions of the music production process. One algorithm from the field referred to as ``reverse engineering a multitrack mix'' can recover the audio effects processing used to transform a multitrack recording into a mixdown in the absence of information about how the mixdown was achieved. This thesis improves on this method of reverse engineering a mix by leveraging recent advancements in machine learning for audio. Using the differentiable digital signal processing paradigm, greybox modules for gain, panning, equalisation, artificial reverberation, memoryless waveshaping distortion, and dynamic range compression are presented. These modules are then connected in a mixing chain and are optimized to learn the effects used in a given mixdown. Both objective and perceptual metrics are presented to measure the performance of these various modules in isolation and within a full mixing chain. Ultimately a fully differentiable mixing chain is presented that outperforms previously proposed methods to reverse engineer a mix. Directions for future work are proposed to improve characterization of multitrack mixing behaviours.
Authors
Colonel, JCollections
- Theses [4223]