Optimal arithmetic dimension of adaptive filters for speech characterization

This paper focuses its scope in the study of the arithmetic accuracy of the results obtained from a real-time implementation of a Speech Spectral Estimator based on an Adaptive Gradient Filter. The arithmetic structure of the filter is implemented in VHDL for a fast synthesis approach to a VLSI ASIC by means of automatic synthesis tools. The arithmetic used is based on 2´s complement fixed-point formats with variable size. The algorithm requires the evaluation of products, additions and divisions. The accuracy of the results are checked vs. standard floating point formats. The results show the dependency of the error In processing different speech signals with the data formats. It is shown that several formats may be advantageously used to reduce the arithmetic complexity preserving numeric accuracy