SIGNAL TRANSFORMATIONS FOR IMPROVING INFORMATION REPRESENTATION, FEATURE EXTRACTION AND SOURCE SEPARATION

This thesis is about new methods of signal representation in time-frequency domain, so that required information is rendered as explicit dimensions in a new space. In particular two transformations are presented: Bivariate Mixture Space and Spectro-Temporal Structure-Field.
The former transform aims at highlighting latent components of a bivariate signal based on the behaviour of each frequency base (e.g. for source separation purposes), whereas the latter aims at folding neighbourhood information of each point of a R^2 function into a vector, so as to describe some topological properties of the function. In the audio signal processing domain, the Bivariate Mixture Space can be interpreted as a way to investigate the stereophonic space for source separation and Music Information Retrieval tasks, whereas the Spectro-Temporal Structure-Field can be used to inspect spectro-temporal dimension (segregate pitched vs. percussive sounds or track pitch modulations). These transformations are investigated and tested against state-of-the-art techniques in fields such as source separation, information retrieval and data visualization.
In the field of sound and music computing, these techniques aim at improving the frequency domain representation of signals such that the exploration of the spectrum can be achieved also in alternative spaces like the stereophonic panorama or a virtual percussive vs. pitched dimension.