Exploiting Temporal Correlation of Speech and Noise Magnitudes Using a Modified Kalman Filter for Speech Enhancement

Abstract:
A new speech enhancement algorithm using a modified Kalman filter in the frequency domain is proposed. The new approach consists of two steps. In the first step, the temporal trajectories of the speech and noise magnitudes are modeled by low order autoregressive (AR) processes, i.e., the current coefficients are propagated in time based on information taken from previous, enhanced coefficients, followed by a subsequent phase estimation. In the second step, the first estimation is updated. Therefore, two statistical estimators are utilized. The performance of the proposed method is shown to be considerably better than purely statistical estimators.