Neural Audio Sculpting Towards Autonomous Multitrack Sonic Transformations

Abstract

This paper explores the nascent application of deep neural networks to the intricate process of multitrack audio manipulation, aiming to redefine traditional mixing paradigms. We survey existing intelligent audio production systems and highlight the emerging integration of deep learning techniques, particularly in content-based audio transformations. A key proposition is a research trajectory focused on leveraging deep learning for intelligent music production. As a foundational proof of concept, we present a deep autoencoder architecture designed to learn and apply implicit audio effect chains to individual stems based on raw input and processed target frequency content. Preliminary results demonstrate the network’s capacity to retain core harmonic and envelope characteristics, despite introducing some artifacts. This work lays the groundwork for future systems capable of autonomous or assistive sonic sculpting in complex audio environments.