Blind separation method of acoustic signals from convolution mixture thereof

摘要

The blind separation method of acoustic signals si1 (t) through sin (t) from a convolution mixture thereof according to the present invention is characterized in that in the first step there is determined a main relative delay diij, wherein i represents i-th microphone and has the value i=2 through m, and j denotes j-th signal and has a value j=1 through n, constituting main waves of acoustic signals coming from acoustic signal sources (2, 20) to microphones (3, 30) relative to a reference microphone (3, 30). In the second step, there is then composed a data matrix X, the lines of which comprise N signal samples xik (t), wherein k represents a microphone (3, 30) index and amounts to k=1 through m from the microphones (3, 30), whereby owing to the main relative delays diij, these samples are shifted relative to each other by a main shift piij of the samples, wherein i represents the i-th microphone (3, 30) and has the value i=2 through m and j denotes the j-th signal and has the value j=1 through n. At least two lines are available in the data matrix X for each acoustic signal si1 (t) through sin (t), wherefrom the main wave of that acoustic signal si1 (t) through sin (t) has identical main relative delay diij, wherein subsequently L-1 lines are further assigned to each line representing the shift piij+1 through piij+L-1, which correspond to reflected waves of the acoustic signals si1 (t) through sin (t), wherein no duplicate lines are present in the data matrix X. Further, algorithm for analysis of independent components is applied to the data matrix X so created to thereby creating a square matrix W. Through subsequent matrix multiplication of said square matrix W with the data matrix X, there is prepared a matrix C comprising statistically independent components of acoustic signals si1 (t) through sin (t). In the next step, affinity matrix D of the individual independent components of the acoustic signals si1 (t) through sin (t) is calculated through the mediation of projection operators, and subsequently clusters of the independent components of the acoustic signals si1 (t) through sin (t) are created by a clustering algorithm according to the affinity matrix D. In the fourth step, a reconstructed data matrix X is created from the independent components of the individual clusters for each acoustic signal (t) through sin (t). Responses of a separated signal corresponding to a given cluster of the independent components are then created from said reconstructed data matrix X on all the m microphones (3, 30). Subsequently, responses of each separated signal of all the m microphones (3, 30) are combined into a single channel to thereby obtaining separated acoustic signals si1 (t) through sin (t).