Abstract

This paper presents an efficient approach to adaptive system identification in undermodelled cases, ie. when the impulse response to be identified is too long to be represented by an adaptive FIR filter operating within the computational benchmarks of a DSP chip. Beyond computational savings gained from adaptive filtering in decimated subbands, the subband approach enables to use different filter lengths in the frequency bands through distribution of taps from an overall number of filter weights determined by the computational benchmark. Compared to a fullband filter of the same complexity, this method can greatly enhance the model representation of long impulse responses. 1 Introduction When identifying very long impulse responses with an adaptive filter implemented on a DSP, as eg. found in acoustic echo control, often the benchmark limits the identification to an undermodelled problem, where IIR filters are not applicable to the nature of the problem [6]. Two possibilities to reduc...

Citations