| Summary: | 碩士 === 國立交通大學 === 電信研究所 === 85 === The normalized LMS is widely used because of its simplicity.
However, the algorithm behaves unstably when all
elements of an input vector is very small with respect
to noise power. Coefficients of the
adaptive filter
are not updated when a norm of an input vector is smaller than a
threshold. But we proposed an improved method that
coefficients are updated by a smaller stepsize if this
condition happens.
Due to the long reverberation time of a room, it is necessary to
use filters of several thousands taps.This result causes slow
convergence speed and high computational complexity.
In the thesis, a subband NLMS based on wavelet packets can
overcome the difficulty. Because of the downsampling
process causes aliasing, adjacent band adaptive filters are
proposed in order to eliminate it. But adjacent band
filters result in slow convergence, we come up with the improved
refinement iteration to speed up convergence speed.
|
|---|
