| As voice communication becomes an ever-more important and pervasive part of our everyday lives, the issue of speech quality becomes more critical. One of the reasons for the undesirable quality degradation is the appearance of audible echoes. This kind of quality degradation is inherently from network equipment and end-user devices. To increase speech quality and improve listening experience, it is necessary to design effective acoustic echo cancellation systems. Echo cancellation has been studied for several decades, and today it is easy to implement echo cancellers on digital signal processors (DSPs). However, certain difficulties still remain to meet the requirements imposed by the echo cancellation standard, and some fundamental challenges still wait for breakthroughs. One of them is the nonlinearity in the acoustic echo path. Nonlinearity usually comes from the price competition in the market of consumer electronics. For economic purposes, the small-sized and low-cost analog components that exhibit nonlinearity, such as loudspeakers and power amplifiers (PAs), are utilized. An echo canceller performs poorly or does not work at all in the system where the net nonlinear distortion is higher than a certain value. In this dissertation, we address the aforementioned nonlinearity issue in acoustic echo cancellation systems. To sufficiently remove the nonlinear acoustic echo, nonlinear adaptive filters have been proposed in the literature to identify the nonlinear acoustic echo path. The identification is done by minimizing the mean square error (MSE) between the microphone-received signal and estimated echo signal. In this way, the echo signal can be reconstructed and subtracted from the microphone-received signal. However, the issues of stability, convergence rate, and computational complexity inhibit nonlinear acoustic echo cancellers (NAECs) from practical implementation. Thus, we are motivated to design efficient NAECs in terms of stability, fast convergence rate, and low computational complexity. First, we propose to perform nonlinearity identification based on the coherence function, which guarantees the stability of the nonlinear adaptive system. Later on, we present a general framework for echo cancellation systems using a shortening filter that entails low computational burden and fast convergence rate. Moreover, we develop methods to remove the system nonlinearity based on the coherence function, including the predistortion linearization, nonlinear residual echo suppressor, and Hammerstein-Wiener model-based NAEC. To design an effective AEC is more than performing an system identification. Another important issue for an AEC is the control logic design of filter adaptation. This problem is caused by the interference at the near-end, including ambient noise and double-talk, when both the far-end and near-end talkers speak at the same time. When double-talk occurs, the adaptive filter may not converge and the identification of the echo path becomes difficult. Double-talk detectors (DTDs) can be utilized to detect the presence of the near-end speech and halt the AEC adaptation, thus to avoid filter divergence. However, DTD designs can be quite complicated since it is often not easy to discriminate between the echo signal and the near-end speech. Moreover, to the best of our knowledge, DTD has not been proposed in conjunction with nonlinear AECs. Unlike double-talk, ambient noise of persistent existence. Therefore, filter adaptation rate needs to be continuously adjusted according to the noise characteristics, rather than being controlled based on carrying out detection. However, few of the learning-rate control algorithms are designed specifically for acoustic echo cancellation applications, which results in the ineffectiveness of these approaches in echo cancellation systems. In the second part of this dissertation, we focus on the control logic design issue. For double-talk detection, we propose to design a DTD based on the mutual information (MI). We show that the advantage of the MI-based method, when compared with the existing methods, is that it is applicable to both the linear and nonlinear scenarios. Furthermore, we extend the MI-based DTD design to the stereophonic acoustic echo cancellation systems. For learning-rate adjustment, we propose a variable step-size and variable tap-length LMS algorithm. Based on the fact that the room impulse response usually exhibits an exponential decay envelop in acoustic echo cancellation applications, the proposed method finds the optimal step size and tap length at each iteration. Thus, it achieves faster convergence rate and better steady-state performance. |
