English

Blind Source Separation

Blind Signal Processing is now one of the emerging areas in Signal Processing with many potential applications like e.g. speech enhancement, audio source separation, etc.

The objective in a blind signal processing problem is to find an inverse system and to estimate the primary source signals for an unknown multiple-input (sources) / multiple-output (sensors) system.

The case where more sources than sensors are available is called underdetermined or overcomplete. Binaural hearing corresponds to this case: Humans are adept at listening to one voice in the midst of other conversations and noise. This attention ability has been termed cocktail party effect.

Our technology demonstration program performs channel separation of stereo signals into center (e.g. lead vocals) and left/right channels. Note that for this application the mixing matrix may be assumed to be known a priori but the separation problem is still overcomplete.