Application of the tuned Kalman filter in speech enhancement

Abstract

The Kalman filter has a wide range of applications, noise removal from corrupted speech being one of them. The filter performance is subject to the accurate tuning of its parameters, namely the process noise covariance, Q, and the measurement noise covariance, R. In this paper, the Kalman filter has been tuned to get a suitable value of Q by defining the robustness and sensitivity metrics, and then applied on noisy speech signals. The Kalman gain is another factor that greatly affects filter performance. The speech signal has been frame-wise decomposed into silent and voiced zones, and the Kalman gain has been adjusted according to this distinction to get best overall filter performance. Finally, the algorithm has been applied to clean a noise corrupted known signal from the NOIZEUS database. It is observed that significant noise removal has been achieved, both audibly and from the spectrograms of noisy and processed signals.