On the application of a psychoacoustically motivated speech-quality measure in CELP speech-coding

The crucial task in a CELP speech codec consists of finding the optimal excitation vector for the synthesis filter. This is usually done in an ‘analysis-by-synthesis’ structure by minimizing the mean-square error of the original and the coded/decoded speech frame. It is a common assumption that distance measures other than MSE and adapted to the human auditory perception should result in better speech quality. Such measures could be based on scientific results provided by psychoacoustics. However, due to the computational load there is no possibility to implement complex psychoacoustical models in real-time speech codecs and, for the time being, we are restricted to the MSE. Nevertheless, it is interesting to study the potential of psychoacoustic distance measures to improve speech codecs if complexity restrictions are neglected. This paper shows how a psychoacoustics-based distance measure can be integrated into a CELP codec, and the unexpected results are presented.