A robust adaptive speech enhancement system

The paper proposes a novel adaptive speech enhancement system with adding random noise (ARNANC) that is little sensitive to the leakage from the primary speech signal into the noise reference signal. The ARNANC speech enhancement system is accomplished by adding a low-level, broadband random training signal to the noise reference signal, and adaptive modeling the transfer function of the noise (NTF) by taking the training signal as the reference signal, eliminating the speech signal interference that will affect the convergence of the modeling filter by using an adaptive prediction filter (APF), modifying the distortion of the training signal component due to the APF by using a compensation filter (CPF). The computer simulations confirm the ARNANC speech enhancement system can effectively separate the primary speech signal from the noisy speech whether there exists leakage from the primary speech signal into the reference input or not.