Improved Audio Mosaicing Techniques for Sound Synthesis

Abstract

An audio mosaic is a piece of audio generated algorithmically by resynthesizing a target sound from one or many source sounds. There has been much research into non-negative matrix factorization algorithms for separation of sound sources from mixed audio signals. These have been successful in breaking down an audio signal's spectrogram into a parts-based representation. We propose an algorithm which uses a non-negative matrix factorization algorithm to separate a source audio le into its latent components, then uses those components to resynthesize another piece of audio through tting by non-negative matrix factorization. We discuss techniques for improving the quality of our audio mosaics, and we provide an evaluation of these techniques on synthesizing audio mosaics from recorded music.