Algorithms for acoustic echo cancellation in the presence of double talk.

Acoustic echo cancellation is an attractive way to enhance speech quality and provide a full-duplex communication channel for hands-free telephony. Among the many implementation challenges, the problem of "double talk" is considered one of the toughest. When near-end speech and the acoustic echo of the far-end speech are simultaneously present in the microphone signal, it is very difficult to separate these two components and provide satisfactory echo attenuation. After providing a thorough overview of the application and the major adaptive filtering algorithms used in echo cancellation, this thesis presents an in-depth study of the published techniques for dealing with double talk. We classify these existing approaches according to their underlying methodology, such as signal level comparison, signal correlation, weight progression monitoring, and dual filter architecture. The strengths and weaknesses of a number of representative algorithms are discussed in some detail. We argue that detection of echo path changes yields a more robust echo canceller than direct double talk detection. The major contribution of this thesis is the development of a novel variable step size approach that uses a measure of the correlation between successive gradient estimates to control the system's adaptation rate. Two main algorithms, denoted GC-VSS3 and PC-VSS, are derived by applying this approach to the Least Mean Squares and Affine Projection algorithms, respectively. We demonstrate the superior performance of the proposed algorithms in a double talk environment. This is done through a series of simulation experiments using white noise, coloured noise, and recorded speech as the input signals. We also show how an auxiliary double talk detector, based on the proposed variable step size technique, may be employed to control the peripheral echo suppression components that are usually required in a hands-free telephone set. Finally, computational complexity is analyzed and shown to be reasonable for practical implementation when certain beneficial properties of the adaptive filter structure are exploited