Speech enhancement by Kalman filtering with residual noise clipping

In this paper, we develop an improved speech enhancement system based on Kalman filtering (KF) that processes the linear prediction (LP) residual of voiced speech. For input frames where voiced speech is present, the LP residuals are clipped to maintain the peaks where the major excitation pulses are located. The proposed algorithm differs from the conventional Kalman filtering algorithms in that it includes this quasi-periodic term in the process equation for voiced speech frames. The quality of the resulting enhanced speech is evaluated by means of signal-to-noise-ratio (SNR) and ITU-PESQ scores. Experimental results indicate that the proposed algorithm achieves consistent improvement in output speech quality when compared to conventional KF methods.