Signal Separation in audio environments have functionality in the area
of Speech Enhancement (i.e. front end to speech recognition unit) and
applications where the enhancement of the entire audio scene is
required (i.e. hearing aids and intelligent devices).
Existing separation techniques can be categorised into two distinct
approaches, model based approaches that exploit high level knowledge
of signals and data-driven approaches that work on the data
directly. These techniques make assumptions and require prior
knowledge that limit application to a mixed audio environment. It is
the intention of this research to develop a separation framework that
is more suitable for audio scene application, through blending
together aspects of the model and data based approaches.
My presentation will begin with an introduction to the field of signal
separation and the objectives of my research. I will then address some
of the limitations of existing separation techniques in an audio
scene. A high level description of an architecture that blends signal
modeling and statistical optimisation will be presented. It is hoped
that this architecture will be appropriate to guide the direction of
my research. I will conclude by discussing some of my research on
Mutual Information.
