ON-LINE BLIND SEPARATION OF MOVING SOUND SOURCES
Jörn Anemüller, Tino Gramß · 1999
In this paper, we propose a method for the on-line blind separation of sound sources in the case where the mixing filters have aÆ-shaped impulse response. Our algorithm works entirely in the frequency domain and exhibits fast convergence due to cross-frequency couplings. Specific problems related to on-line separation of running speech are discussed. The algorithm performs successful separation of digitally mixed speech signals and of signals from a moving and a standing speaker recorded in an anechoic chamber. Ñ�Ø � �����×�Ø � (1) ofÆindependent source signals×�Øresulting inÆsensor signalsÑ�Ø, various algorithms for blind source separation have been proposed; refer to, e.g., [4]. Their com-�ÜØ��Ü as�ÜØ mon goal is to find an estimateÏ��Ï��℄for the inverse of the mixing matrix������℄and reconstruct the sources �Ï�ÑØ. (We use the notation Ø�����ÜÆØ℄Ì, etc.) The source separation algorithm used by us is derived through the maximum-likelihood method [3, 6], augmented for nents×�asÈ×��È��×��� complex-valued signals. Since we are dealing with Fourier transformations of speech Ó×� signals, we model the probability densities of the corresponding spectral compo-(2) 1. SOURCE SEPARATION In the case of an instantaneous linear superposition for the weight-matrixÏat time-stepØ. The transposition and complex conjugation of�Üis denoted as�ÜÀ. We denote by�the adaptation step’s size. In the remainder of this paper, two well-known symmetries of source separation algorithms are relevant: scalinginvariance, i.e., the sources can only be reconstructed up to a constant scaling factor, and permutation-invariance, i.e., it is unknown which reconstructed signal resembles which source. 2. ACOUSTIC SOURCE SEPARATION IN THE