Design of a trellis enumeration system for the scalar quantifier of the Federal Standard FS 1016

Abstract

In the present study, we were interested in the reduction of the bit-rate observed in the speech coder named CELP FS1016 (federal standard developed by the US department of the defense " DoD "). More precisely, the quantization of the Line Spectrum Frequencies (LSF) parameters was concerned. In the standard CELP FS1016, these coefficients are derived from the input speech signal through linear prediction analysis of each 30-ms frame. A direct scalar quantization at 34 bits per frame is used. We considered the bit rate too high and we propose the use of an enumeration technique in conjunction with a treillis search coding schemes for speech LSF parameters, to reduce this rate to 30 bits per frame. To this end. we first did a thorough analysis of the different possible combinations given by the tables of the scalar quantization initially used in the FS 1016. The analysis revealed that this quantization was not optimum. Indeed, we found that many combinations of quantization levels can not be used because they can lead to an unstable synthesis filter. Then, we exploit the natural ordering of the LSF to design an enumeration algorithm with a trellis schemes that reduces the bit rate of the LSF coefficients from 34 to 30 bits without decreasing the performance of the coder (an equivalent spectral distortion is obtained with the two quantizers).