Performance of a fast algorithm for FIR system identification using least-squares analysis

A wide variety of procedures have been proposed for identifying a finite impulse response (FIR) linear system from the input and output of the system. Most recently, a fast, efficient, least-squares method was proposed by Marple, and was shown to require less computation and storage than any other known procedure for identifying moderate to large FIR systems. In this paper we measure the actual performance of the newly proposed fast system identification algorithm by using it to estimate a variety of FIR systems excited by either white noise or a speech signal. It is shown that essentially theoretically ideal performance is achieved for white noise inputs; however, for speech signals poor performance was obtained because of the lack of certain frequency bands in the excitation. A simple modification to the estimation procedure is proposed and is shown to provide substantial performance improvements. Using the spectrally modified speech signal, the performance of the fast system identification algorithm was found to be acceptable for a wide variety of applications.