Blind signal separation:"Cocktail party problem"

jkormanski said:

Hello, i am working currently on algorithms for solving "cocktail party problem". I am interested if anyone works on this topic. I would like to exchange some knowledge about this topic.

Click to expand...

jkormanski said:

"Cocktal party problem" can be described as a situation when we have 2 (or more) speakers talking simultaneously. Each speaker is a source and we record a mixture of the sources using 2 microphones (sensors). Each sensor contains a mixture of voices of 2 speakers. It`s a 2x2 MIMO system. The task is to recover the individual sources(voices) having only the record of their mixtures (the mixing process of sources is unknown).
By this time i wrote an algorithm in matlab which solves this problem.The recover(demixing) process is done with natural gradient adaptation algorithm.
When i tested the performance of demixing algorithm i took 2 wav files,one with the voice of my brother and second with my own voice and mixed them together. The mixing was done with a mixing 2x2 matrix by multiplying matrix by source vectors to give 2 vectors of mixed speech ( 2 sensors). This system is invertible so if we have the observations of signals (records from sensors) then if the mixing matrix is known, original sources can be found by multiplying the inverse mixing matrix by sensor vectors. Of course in blind signal separation tasks we assume that the mixing process is unknown (we dont know the values of the mixing matrix) and we want to recover original sources from their mixtures.So it can be said that problem will be solved when the demixing matrix will be found.
I hope that the whole problem and solution is clear. In my research i want to compare different kinds of demixing algorithms. The next algorithm i want to concentrate on is a demixing algorithm using genetic algorithms beacuse they perform well in finding optimal solutions.
I also plan to model some acoustic path( like a room impulse response) when mixing sources: adding echo effect which will affect that mixture will be with an echo effect and then recover orginal voices without echo.
If someone is interested i can put my m-files and also some material coevering this topic in here.

Click to expand...