Analysis of signal separation and distortion analysis in feedforward blind source separation for convolutive mixture

This paper analyzes signal separation and distortion in a feedforward blind source separation (BSS) applied to convolutive mixture environment. In the BSS learning, a separation block is adjusted so as to make the output signals statistically independent. One direction for this adjustment is to extract only one signal source at a single output. The other direction is to make the output signal spectra white. The second direction causes signal distortion and conflicts with the first direction. This confliction prevent signal separation. These relations are also dependent on a length of the impulse responses in the mixture and the filter length in the separation block. These properties are theoretically analyzed and are investigated through simulations using the white signal sources and the speech signal sources.