GMU Microsound Analysis

In 1946, Dennis Gabor introduced the idea that any sound could be decomposed into a set of simple acoustical events called quantum. This granular model of sound was perceptually validated according to the limitations of the auditory system dealing with time-frequency discrimination. Each of these acoustical events corresponds to a local Time-Frequency component of the sound and thus can represent a large variety of signal structures from transients to pitched sustained parts.

This theory of signal and perception introduced a new approach to synthesize sounds called Granular Synthesis. The main principle of this technique is the accumulation of a large amount of basic parametric sonic events called grains. Iannis Xenakis was one of the first music composers who used the grain as the basic symbolic component of some of his pieces (mostly instrumental) and thus broke the wall between micro and macro musical structure. Since the 1970s, Curtis Roads has explored many aspects and applications of this synthesis technique from real-time pitch shifting of sounds to complex textures generation. He particularly studied the perception effects of the different synthesis parameters and proposed an exhaustive categorization of the diverse applications according to the constraints/relations applied to these parameters. Thus, lots of high-level control strategies have been introduced but with an empirical background. From here comes our interest to deduce granular synthesis parameters from previously analyzed sound, that is to say to design a granular analysis/synthesis tool.

This idea comes with the application of such techniques to natural noisy sounds in mind. It points to sounds that can be defined as the accumulation of more or less complex sonic grains, with their proper temporal and spectral variability. For example, the sound of rice falling onto a metal plate is composed of thousands of elementary “ticks”; the rain produces, in the same way, the accumulation of a large amount of water droplet microsounds...
In fact, in the real world, when multiple realizations of a same event, of a same phenomenon occur, we can expect these types of sounds. Our goal is thus to analyze natural sounds in order to extract the temporal and spectral distribution of those grain streams, and to model these evolutions/fluctuations by correlated statistical laws. We expect, in fine, the possible synthesis of sounds perceptually ascribable to the class of the analyzed sound. This might lead us to the classification of granular sounds according to their characteristics. Thus we have to detect the grain parameters of the analyzed sound. In this purpose, we experimented several techniques of sound decomposition based on the Matching Pursuit algorithm. We also proposed an extension of this algorithm to the spectral domain that we have called Spectral Matching Pursuit.

For more info, see ICMC paper (in English) and JJCAAS poster (in French)