Harmonics Enhancement for Determined Blind Sources Separation using Source’s Excitation Characteristics

We present an improved method on combining temporal and spectral processing approaches for multichannel determined blind sources separation. The separation task is performed by applying the spectral processing on a mixed speech, using sources’ excitation characteristics. The performance of the proposed method is investigated by separating two sources from a stereo recording mixture extracted from BSS-Locate [1]. Evaluation is performed by objective quality measure BSS-eval tool [2], perceptual evaluation of speech quality (PESQ), and Short-time Objective Intelligibility Measure (STOI) [3]. Simulations allow comparison with an existing spectral processing approach (TSP), and clearly demonstrate the efficiency and the outperformance of the proposed method. Keywords— Speech separation; LP residual; Glottal Closure Instants; time delay of arrival; Hilbert Envelop