SEPARATION OF NON ORTHOGONAL SPECTRAL DATA

Danielle Nuzillard
e­mail: danielle.nuzillard@univ­reims.fr

Independant component analysis relies on the sta­ tistical independance of the sources, a constraint that is not always fulølled when dealing with particular real life problems. The positivity of mixing coeOEcients and of spectral source data, imposed by physical reasons, is also a strong constraint that permits to improve the solution of source separation problems. The neces­ sity of dealing with spectral data led ørst to adapt the Second­Order Blind Identiøcation (SOBI) algorithm to frequency domain data sets. The SOBI algorithm is not able to retrieve non­orthogonal sources from mix­ tures. However, it may produce solutions that are close to reality. ATheir reønement through the Alternated Least Squares procedure (ALS) introduces the positiv­ ity constraint and improves greatly the quality of the separation.