UNVEILING THE TIMBRE LANDSCAPE: A LAYERED ANALYSIS OF TENOR SAXOPHONE IN RAVE MODELS

Abstract

This paper presents a comprehensive investigation into the explainability and creative affordances derived from navigating a latent space generated by Realtime Audio Variational AutoEncoder (RAVE) models. We delve into the intricate layers of the RAVE model's encoder and decoder outputs by leveraging a novel timbre latent space that captures micro-timbral variations from a wide range of saxophone extended techniques. Our analysis dissects each layer's output independently, shedding light on the distinct transformations and representations occurring at different stages of the encoding and decoding processes and their sensitivity to a spectrum of low-to-high-level musical attributes. Remarkably, our findings reveal consistent patterns across various models, with the first layer consistently capturing changes in dynamics while remaining insensitive to pitch or register alterations. By meticulously examining and comparing layer outputs, we elucidate the underlying mechanisms governing saxophone timbre representation within the RAVE framework. These insights not only deepen our understanding of neural network behavior but also offer valuable contributions to the broader fields of music informatics and audio signal processing, ultimately enhancing the degree of transparency and control in co-creative practices within deep learning music frameworks. (c) 2024. This is an open-access article distributed under the terms of the Creative Commons Attribution 3.0 Unported License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original.