| Eng | Gr |

Nikolaos Mitianoudis

Professor, Democritus University of Thrace

Cocktail Party Problem

The problem of Blind Source Separation is more clearly illustrated by the coctail party problem. Imagine the situation of being in a coctail party with a lot of people talking simultaneously. The coctail party problem is defined as the problem of using your ears (sensors) to isolate every speaker (source) in the room while suppressing the others. This problem can be extended to the M sensors, N sources case, leading to the general Blind Source Separation case.

Audio Source Separation using Independent Component Analysis

Many methods have been applied to solve BSS. Recently, we have seen the emergence of Independent Component Analysis, that proved to be extremely efficient in the separation of instaneous mixtures (a very special case of BSS). Our basic research effort is to apply ICA techniques to some other interesting source separation cases (i.e. sources recorded in a real room environment), using fast Newton-type ("fixed-point algorithms"), and extend these techniques to musical instrument separation from audio recordings. Other aspects involved in my project are: ICA using multiresolution analysis, ICA of more sources than sensors and ICA in the presence of noise.

Blind Source Separation demo

You can listen to some audio source separation demos of the algorithms presented in the following papers:

• " Audio source separation of convolutive mixtures", IEEE Transactions on Speech and Audio processing, Volume: 11, issue: 5, pages 489-497, September 2003.
• "New fixed-point solutions for convolved mixtures", 3rd International Conference on Independent Component Analysis and Source Separation, San Diego, California, December 2001
• "A fixed point solution for convolved audio source separation", IEEE workshop on Applications of Signal Processing on Audio and Acoustics, New Paltz, New York, October 2001.

Underdetermined source separation using Directional Laplacian Mixture Models

In this project, the Blind Source Separation problem of overcomplete instantaneous mixtures is examined, in the case that the source signals are very sparse. A Generalised Directional Laplacian Mixture Model (MDLD) is trained using the Expectation-Maximisation (EM) algorithm to separate sources existing in multidimensional mixture setup. The individual Laplacians of the mixture model can help identify the mixing matrix and the source signals using optimal detection theory. A hard and a soft decision scheme were introduced to perform separation.

Some samples demonstrating the algorithms and the experiments proposed and conducted in following paper are presented here:

• N. Mitianoudis, "A Generalised Directional Laplacian Distribution: Estimation, Mixture Models and Audio Source Separation", IEEE Transactions on Audio, Speech and Language Processing, Vol. 20, No. 9, Nov. 2012.