A Controlled Noise Reduction Wiener Filter Based on the Quadratic Eigenvalue Problem

This paper deals with the problem of multichannel noise reduction and a controlled Wiener filter is developed in the frequency domain by minimizing the subband mean-squared error between the clean speech of interest and its estimate subject to a constraint on the residual noise level. Using the Lagrange multiplier, we then transform the constrained minimization problem into one of quadratic eigenvalue problem. Different forms of the controlled Wiener filter are then derived. Depending on how the quadratic eigenvalues are used, these filters can control the amount of noise reduction as well as the compromise between the amount of noise reduction and the level of speech distortion. Simulation results justify the properties of the controlled Wiener filter.