Stochastic analysis/synthesis using sinusoidal atoms

This work proposes a method for re-synthesizing music for use in perceptual experiments regarding structural changes and in music creation. Atoms are estimated from music audio, modelled in a stochastic model, and re-synthesized from the model pa- rameters. The atoms are found by splitting sinusoids into short segments, and modelled into amplitude and envelope shape, frequency, time and duration. A simple model for creating envelopes with percussive, sustained or crescendo shape is presented. Single variable and joint probability density functions are created from the atom parameters and used to re-create sounds with the same distribution of the atoms parameters. A novel method for visualization music, the musigram, permits a better understanding of the re- synthesized sounds.