Co-channel speech separation based on adaptive decorrelation filtering

The problem of co-channel speech separation based on adaptive decorrelation filtering (ADF) is studied. The ADF algorithm proposed by Weinstein et al. is carefully examined as the foundation of this work. The algorithm is formulated for the separation of two co-channel speech signals acquired by two microphones. By modeling the underlying acoustic paths using linear FIR filters and adaptively estimating these filter coefficients from the acquired signals based on decorrelation, the speech signals generated by different sources can be separated through a filtering process using the estimated filters. For practical application of this algorithm, improvements are proposed to address the issues of computational complexity, system stability, and convergence performance. The limitation of the algorithm is also discussed to identify suitable application scenarios.; The two-source ADF algorithm is generalized for the separation of M > 2 speech sources to extend the scope of application. A modification on the generalized algorithm is also proposed for scenarios where some speech sources are not of interest. By not extracting these undesired speech sources, this modification results in lower computational complexity, improved dynamic tracking ability, and better steady-state separation performance. An analysis on robustness shows that the ADF algorithm is sensitive to the existence of background noise, especially when the noise components in the acquired signals are highly correlated. It is also shown that by integrating a speech enhancement front-end based on the energy-constrained signal subspace method, the robustness of the ADF algorithm against non-speech-like background noise can be significantly improved.; Finally, to further improve the separation performance achieved by the previous direct-form ADF as well as to facilitate modular and easy implementation of ADF in hardware, a lattice-ladder structure for ADF is developed based on the joint forward and backward linear predictions. Experimental results demonstrate the effectiveness of the lattice-ladder algorithm in reducing cross-interference between co-channel speech sources. The results also show that the lattice-ladder algorithm achieved significant performance improvement over the direct-form algorithm. A simplified lattice-ladder ADF is proposed as a compromise between computational cost and system performance.