Transform domain prediction error method for improved acoustic echo and feedback cancellation

The prediction error method (PEM) has been successfully applied in double-talk-robust acoustic echo cancellation (AEC) as well as in acoustic feedback cancellation (AFC). The main idea in both applications basically consists in decorrelating the adaptive filter input and error signals. This is done by whitening these signals with the inverse of a near-end signal model before the filter adaptation. The choice of the near-end model greatly affects the performance and complexity of the resulting AFC/AEC algorithms, oftentimes turning the algorithm impractical for real-world realtime applications. This paper proposes the use of discrete cosine transform (DCT), in conjunction with a simple near-end signal model, to boost the performance of PEM-based algorithms both in double-talk-robust AEC and AFC while only marginally increasing the computational complexity.