A new algorithm on multiple unknown source signals estimation and separation in a reverberant space (Q1970912)

The aim of the paper is to propose and investigate a new algorithm, based on a finite impulse response (FIR) model, in order to perform the estimation and the separation of signal sources in a noisy reverberant environment without knowing the actual number of sources. The channel parameter matrices are estimated by using the statistical properties of a pseudo-white noise as the reference signal. The accuracy of the estimated channel characteristic matrices depends on the covariance of the reference signals, but it can be controlled manually during the implementation procedure. The estimation and the separation of the source signals is achieved on the basis of a recurrence formula. The advantages of the proposed algorithm is that both the signal waveforms and their amplitudes from the original source signals can be preserved, and no restrictions are imposed on the source signals, i.e. they can be correlated signals. Furthermore, the estimation and the separation of the source signals are performed in different stages, which can reduce the computational effort. The validity of the proposed algorithm is examined and compared to the EFOBI (extended fourth-order blind identification) approach. The comparison with the EFOBI algorithm is shown to provide better results for the proposed system, thus to represent a good alternative to the currently existing on-line applications.