D. Arfib
CNRS-LMA, 31 chemin J. Aiguier, 13402 Marseille Cedex 09, France
Ph. Guillemain
DIGILOG, 13852 Aix-en-Provence Cedex 3, France
R. Kronland-Martinet
CNRS-LMA, 13402 Marseille Cedex 09, France
To play natural sounds on an algorithmic synthesizer, it would be necessary to extract parameters that correspond to a synthesis model from an analysis of a real sound. The aim of this lecture is to show different facets of a time-frequency representation that are very particular in the signal-processing field because they are oriented toward the musical use. For example, psychoacoustic criteria allow one to ``get rid of'' the incertainty principle on time and frequency. Here, only the acoustically relevant parameters are looked for rather than the exact physical data. For that purpose, time-frequency transforms were used, namely the Gabor and the wavelet transforms, and additional work on the time-frequency representations they provide was conducted. For example, the spectral line estimation allows one to separate each elementary component and to estimate the modulation law associated to each one by solving a linear system where the coefficients come from the representation. Special features concerning the sound can be derived from ridges extracted from the phase diagram; this can lead to the estimation of frequency modulation laws in a FM model. The evolution of the centroid of the spectrum can be a good indicator for a synthesis using a wave-shaping model. The lecture will be accompanied by a lot of time-frequency representations showing the importance of the choice of the parameters of the analysis, and the diagrams that one can derivate from them. It will also be accompanied by sound examples.