Dual perceptually constrained unscented Kalman filter for enhancing speech degraded by colored noise

A dual unscented Kalman filter (UKF) is applied to enhance speech signals degraded by colored noise. The speech model is a nonlinear one and is modeled by a neural network. Tire state vector and the neural weight vector are estimated by the dual UKF. The UKF to estimate the state vector is a perceptually constrained UKF. The constraints are derived from the masking properties of human auditory systems. Simulation results show that the algorithm can produce better PESQ scores than some recently published methods.