|
Abstract:
The application of neural network based adaptive filters to nonlinear acoustic echo cancellation is studied in this paper. A two stage structure consisting of a parallel arrangement of a temporal neural network and a conventional transversal filter is constructed. A modified form of the partial conjugate gradient algorithm which uses a gradient average window and a fixed step size is used to train the network. Experimental results indicate the new structure is capable of improving the convergence rate compared with backpropagation. The average echo cancellation of real speech signals when passed through a nonlinear loudspeaker in a handsfree telephone conference environment is improved by approximately 5 dB as compared to an FIR filter trained with the accelerated SFTF algorithm.
|
