Acoustic Echo Control Based on Frequency-domain Stage-wise Regression

Traditional echo control techniques as Partitioned Block Frequency Domain Adaptive Filter (PBFDAF) with stochastic gradient adaptive method usually endure slow convergence and insufficient echo suppression in reverberant room when the echo is speech and the echo path is unstable. An algorithm based on frequency domain stage-wise regression is proposed for acoustic echo control to achieve faster convergence of the system estimation with insignificant bias. Commonly used additional double-talk detector and inter-channel coherence based residual echo suppressor are not needed since short-time coherence analysis is performed in each stage. By further making mild assumptions on the quasi-stationarity of the near-end background noise, both fast convergence and stability of the estimation can be achieved simultaneously with a non-stationarity controlled smoothing factor. Experiments are carried out to show the superiority of the proposed approach in terms of convergence speed and steady state error in distant talking mode in ordinary room environment with various common levels of background noise.