Statistical modeling of noisy
sounds:
spectral density, analysis, musical
transformations, and synthesis
Abstract:
Computers offer new possibilities concerning sound processing.
Existing research essentially concerns the modeling of the
periodic parts of sounds. However, noisy sounds are more and more used
in musical compositions.
Existing models consider sounds, or their noisy parts, as filtered
white noise. Psychoacoustic studies about the perception of the
spectral density of frequencies show that this assumption represents a
restriction. We present a new spectral and statistical model, called
CNSS, based on the thermal noise model. A noisy sound is defined by a
fixed number of sinusoids. The CNSS model proposes new controls
related to the distributions of the frequency and phase of the
sinusoids, and thus provides new musical transformations.
A general analysis method for noisy sounds is proposed. We describe a
first method which characterizes partials of the noisy components in
the spectrum by statistically studying their intensity fluctuations. A
second method defines parts of the temporal representation that
contain transients by studying the distribution of the samples of the
signal. Lastly, we detail a new technique which approximates the
spectral density. The parameters of the CNSS model can then be
estimated.
Then, the synthesis algorithms related to the CNSS model are then
detailed. Moreover, problems due to the window overlap during the
synthesis of random sounds are explained and some solutions are
proposed.
The CNSS model has been implemented and provides real-time synthesis
of noisy sounds. This implementation is planned to be integrated in
the computer-assisted early-learning tool Dolabip, which is
developed at the SCRIME. It is also a platform for exploring sounds
and a sound synthesis for psychoacoustic experiences.
Discipline: Computer-Science
Keywords:
statistical modeling,
stochastic part,
spectral density,
analysis and synthesis of sound,
noises,
real-time additive synthesis.
|