Echo Cancellation: A Novel Adaptive Kalman Filter-Based Scheme

Abstract

A novel adaptive echo cancellation scheme, using an accurate and reliable two-stage identification scheme and an adaptive Kalman filter (KF), is proposed. The novel scheme estimates a desired waveform from the received signal which is corrupted by an undesired echo and noise. It is assumed that the desired waveform and the echo are uncorrelated with each other, and that the reference waveform is highly correlated with the echo. Further, the reference waveform and echo are related by a rational transfer function and are assumed to be measureable. An accurate and reliable model of the echo is identified from the received signal using a proposed novel two-stage identification scheme. This two-stage scheme is used to identify accurately and reliably the model relating the reference and actual echo waveforms using least squares and model order reduction. It is implemented in the frequency domain using waveform segmentation to ensure efficient computation and model reduction, signal stationarity and real-time system implementation. The identified model of the echo is then embodied into the KF which is a minimum-variance estimator that is robust to noise and disturbances and has a zero-mean, white noise residual. The performance of the KF is monitored continuously and its gain updated adaptively. If the filter's residual fails the whiteness test, the model of the echo is then re-identified and the KF adapted accordingly. The proposed scheme was successfully evaluated on simulated and real recorded speech corrupted by noise and echo. This novel scheme can be extended to areas such as beamforming