摘要:
Adaptive discrete-time filters, as used, for example, in echo cancellers, comprise a control circuit for controlling the coefficients for adjusting the discrete-time transversal filter. A prior art manner of adjustment is the LMS algorithm. This algorithm is relatively simple because few computations are necessary but they may lead to poor convergence results with autocorrelated input signals. Further, the OP algorithm is known which leads to good convergence results with auto-correlated input signals but which also requires a substantial amount of computation. According to the invention an algorithm is provided for input signals, more specifically speech signals, which may be modelled with an autoregressive process of the order of p, which algorithm requires the same amount of computation as the LMS algorithm but leads to the same or better convergence results for this type of signals than the OP algorithm.
摘要:
A frequency-domain block-adaptive digital filter (FDAF) having a finite impulse response of length N for filtering a time-domain input signal in accordance with the overlap-save method includes window means (11) for obtaining modifications (B(p;m)) of the 2N frequency-domain weighting factors (W(p;m)) from correlation products (A(p;m)). A known FDAF of this type contains five 2N-points FFT's, two of which are used in the window means (11). By utilizing a special time-domain window function which can be implemented very efficiently in the window means (11) with the aid of a frequency-domain convolution, a FDAF of this type containing only three 2N-point FFT's is obtained whose convergence properties are comparable to those of the known FDAF containing five 2N-point FFT's.