Musically expressive sound textures from generalized audio

Ben Recht; Brian Whitman
DAFx-2003 - London
We present a method of musically expressive synthesis-by-analysis that takes advantage of recent advancements in auditory scene analysis and sound separation algorithms. Our model represents incoming audio as a sub-conceptual model using statistical decorrelation techniques that abstract away individual auditory events, leaving only the gross parameters of the sound– the “eigensound” or generalized spectral template. Using these approaches we present various optimization guidelines and musical enhancements, specifically with regards to the beat and temporal nature of the sounds, with an eye towards real-time effects and synthesis. Our model results in completely novel and pleasing sound textures that can be varied with parameter tuning of the “unmixing” weight matrix.