Fast echo cancellation in a voice-processing system

An efficient adaptive echo cancellation scheme where near-end hybrid echoes are considered the dominant factor in poor DTMF detection and speech recognition in the presence of outbound speech is described. The incorporation of distinct and separate adapt and cancel filters is crucial in attaining significant echo reduction. The adapt filter performs nonreal-time identification of the echo transfer function with buffered transmit and receive data that meet specific criteria. The adaptation implementation is controlled by what has been termed cycle steal; after the required functions have been performed, the remaining real time is utilized for adaptation. After each convergence interval, a windowed set of coefficients is passed from the adapt filter to the cancel filter. The cancel filter performs echo cancellation in real time with the defined coefficients and produces dramatic improvements in measured DTMF detection performance