Extended Data for: Quality-Diversity Search in Sound Generation: Investigating Innovation Engines for Audio Exploration

Data accompanying the article Quality-Diversity Search in Sound Generation: Investigating Innovation Engines for Audio Exploration. The Innovation Engine algorithm is used to evolve sounds, where Quality Diversity search is guided by the YAMNet classifier to discover sounds.
This study draws on the challenges that composers and sound designers face in creating and refining new tools to achieve their musical goals. Utilising evolution- ary processes to promote diversity and foster serendipitous discoveries, we propose to automate the search through uncharted sonic spaces for sound discovery. We argue that such diversity promoting algorithms can bridge a technological gap between the theoretical realisation and practical accessibility of sounds. Specif- ically, in this paper we describe a system for generative sound synthesis using a combination of Quality Diversity (QD) algorithms and a supervised discrimi- native model, inspired by the Innovation Engine algorithm. The study explores different configurations of the generative system and investigates the interplay between the chosen sound synthesis approach and the discriminative model. We further examine the interaction between Compositional Pattern Producing Net- works (CPPNs) and Digital Signal Processing (DSP) graphs, introducing a novel approach with multiple specialized CPPNs for different frequency ranges. This configuration results in simpler CPPN networks while maintaining comparable performance to single-CPPN setups. The research also investigates evolution- ary stepping stones by analyzing goal switches between musical and non-musical contexts, revealing how lineages traverse unlikely paths to current elites. Addi- tionally, we explore the temporal dimension of sound generation by expanding the behavior space from a previous study to include various sound durations, uncovering specialization within temporal niches. The results indicate that a combination of CPPN + Digital Signal Processing (DSP) graphs coupled with Multi-dimensional Archive of Phenotypic Elites (MAP-Elites) and a deep learn- ing classifier can generate a substantial variety of synthetic sounds. Our expanded experiments demonstrate the system’s ability to produce diverse and innovative sound objects across different temporal and contextual dimensions. The study concludes by presenting the generated sound objects through an online explorer and as rendered sound files. Furthermore, in the context of music composition, we present an experimental application that showcases the creative potential of our discovered sounds, highlighting the system’s capacity for generating versatile sonic material across various durations and contexts.