Determined blind source separation using features extraction

Among different Blind Sources Separation (BSS) approaches, little attention has been paid to techniques based on features extraction, as they are assumed to be arbitrary and not reproducible. In this paper we demonstrate that the separation task based on feature extraction could compete with other BSS techniques. Our approach performs sources separation task by exploiting observed mixed speech captured by only two microphones. While previous work combines Temporal and Spectral Processing techniques, which is known as (TSP), our method improves the spectral processing approach by only using the temporal processing parameters. The performance of the proposed method is investigated in the determined context, where two sources are captured by two microphones in reverberant environment. Evaluation is based on the use of objective metrics. Simulation and analysis demonstrate that the proposed method leads to the best compromise between separation and intelligibility, compared to other techniques in the field.