Under-determined blind speech separation via the convolutive transfer function and lp Regularization

![CDATA[Convolutive blind speech separation (CBSS) attempts to recover speech sources from recorded mixtures transmitted by room impulse response filters (RIRs). The existing BSS methods are mainly based on a multiplicative narrowband approximation (MNA) model in the short-time Fourier transform (STFT) domain. However, the MNA model may result in serious system transformation error under the situation of a strongly reverberant environment due to its sensitivity to RIRs with large lengths. Moreover, the solution of source reconstruction is not unique when the system is underdetermined even if the RIRs are given in advance. To tackle these problems, a convolutive narrowband approximation (CNA) model is explored in the proposed approach to constrain the system transformation error. Then an l p(0lt pleq 1) regularization is provided to reconstruct the sparse sources in which p can be dynamically selected to achieve the best performance. The experimental results demonstrate the robustness of the proposed method to room reverberation under various speech separation cases in comparison to conventional methods.]]