Télécom Paris

From Source Localization to Blind Source Separation: an intuitive route to convolutive blind source separation
  • Bjoern Schoelling, (Darmstadt University of Technology)
  • Martin Heckmann, (Honda Research Institute Europe GmbH)
  • Frank Joublin, (Honda Research Institute Europe GmbH)
  • Christian Goerick, (Honda Research Institute Europe GmbH)
  • Microphone arrays and array signal processing
  • Sound enhancement and sound separation
  • Adaptive filtering algorithms and structures for echo and noise control
Get the paper in PDF format
Acrobat Reader (version 5 minimum) is necessary to read this document.

Most algorithms for blind source separation (BSS) of convolutive speech mixtures are derived in a deductive way from abstract statistical principles and exploit a combination of three signal properties, i.e. nongaussianity, nonwhiteness and nonstationarity. In this paper we show how a separation system can be build the opposite, inductive, way using basic speech processing building blocks. The main block and starting point of our derivation is a simple generalized cross correlation based localization systemw with two microphones. The capability of source separation (2 signals and 2 sensors) is added by duplicating the localization structure and adding an inhibition mechanism which suppresses already localized sounds.

©2006 Télécom Paris/TSI
Edition : Télécom Paris -- 2006