Acoustic echo cancellation structures based on perceptual hearing criteria.

Abstract

Hands-free communication techniques are becoming increasingly important in the world of modern telephony. The convenience of hands-free links is invaluable in teleconferencing situations, automobiles and even in personal communications. The loud-speaker and microphone on a hands-free set are inherently coupled acoustically, producing an echo which degrades the quality of conversation over such a link. The conventional acoustic echo cancellation method is gain control which provides a half-duplex communication link. Research into adaptive filters which can provide a full-duplex communication link is intense but transparent transmission of speech over a hands-free link remains an elusive goal. Conventional acoustic echo cancellers are concerned with minimizing an objective measure of the level of echo. This thesis presents a study of two adaptive filtering structures which are based on the goal of providing the best subjective echo cancellation possible. The adaptive filtering algorithms attempt to give the echo the same spectral shape as the absolute hearing threshold curve, in order to minimize the echo's audibility. The filtered-E LMS algorithm, considered first, was found not to perform the kind of frequency weighting desired. A proposed subband adaptive filter with perceptually-based adaptive tap assignment, when trained with natural speech, is capable of providing a noticeable improvement in audible echo cancellation; however, when trained with bandlimited speech, the algorithm's performance was found to be only slightly better or comparable to the fixed tap assignment algorithm in terms of audible echo cancellation. The subband adaptive filtering algorithm is evaluated by means of simple listening tests performed by the author.