One method for maintaining accuracy in implementation of Fast Fourier Transform on fixed point digital signal processors

Abstract

High resolution spectral analysis is frequently used in recently developed complex algorithms for audio signal compression. Long filter banks used in those algorithms are frequently realized using Fast Fourier Transform. Due to its recursive nature, calculation of long Fast Fourier Transforms on fixed point DSP can result in significant accuracy loss. Several methods of maintaining accuracy in implementation of Fast Fourier Transform on fixed point DSPs and analysis of their performance are presented in this paper. Firstly, overview of existing methods is given. Then improved method is proposed, and finally, those methods were compared. Segmental signal-to-noise ratio was used as main criteria for comparison. Requirements for DSP resources (processing complexity, memory, hardware support) to maintain accuracy are also considered. © 2006 by MIPRO. All rights reserved.