Blind deconvolution of speech mixtures based on time-frequency processing and statistical analysis

The problem of blindly separating a convolutive mixture of speech signals is considered. The identification and classification of delay and amplitude parameters is achieved through time-frequency processing based on spectral analysis and statistical methods. A new "closed-form" algorithm is proposed and does not require any iterations with the raw data. By separating digital mixes of real-world recording of speech signals and evaluating the performance under various noisy conditions, we demonstrate that the proposed algorithm is robust and can be applied to speech.