Detail

Statistical Modeling and Synthesis of Intrinsic Structures in Impact Sounds

A struck object produces sound that depends on the way the object vibrates.
This sound is determined by physical properties of the object, such as its
size, geometry, and material, and also by the characteristics of the event,
such as the force and location of impact. It is possible to derive physical
models of impact sounds given the relationship between the physical and
dynamic properties of the object, and the acoustics of the resulting sound.
Models of sounds have proven useful in many fields, such as sound
recognition, identification of events or properties (e.g. material or
length) of the objects involved, sound synthesis, virtual reality and
computer graphics. However, physical models are limited because of the a
priori knowledge they require and because they do not successfully model all
the complexities and variability of real sounds.

In this talk, we propose data-driven methods for learning the intrinsic
features that govern the acoustic structure of impact sounds. The methods
are able to characterize the structures that are common to sounds of the
same type as well as their variability. They require no a priori knowledge
and aim for low dimensional characterizations of the sounds. In addition,
they are not restricted to learn an explicit set of properties of the sounds
(e.g., basic features such as decay rate and average spectra); instead, they
learn the properties that best characterize the statistics of the data. The
methods can learn properties of the sounds such as ringing, resonance,
sustain, decay and sharp onsets.

In this talk, we also explore the synthesis of impact sounds using the
features learned by these methods. We will show that it is possible to
manipulate the learned features in order to modify the original sounds, or
even to create new sounds (for example, from the interpolation of the
representation of recorded sounds). Finally, we will show that the sounds
synthesized by the methods are realistic, as they are perceived more often
as real than as synthesized.