Optimized frame selection for variable frame rate synthesis

In LPC voice output systems, memory requirements can be reduced if the parameter curves are replaced by straight-line trains and only the parameter sets at the segment boundaries are stored. In this paper an optimization algorithm is described, which for a given number of segment boundaries and a given error criterion calculates the optimal positions of the boundaries. In a first application of this algorithm, the original ordinate values of the boundaries were stored and the RMS error between the original and the interpolated LPC parameters was used as optimization criterion. Then the ordinate values were also optimized. In a third experiment an error criterion derived from psychoacoustic parameters, the specific loudness functions, was used. In each case the aim was to determine the minimum number of parameter sets for which no difference to the original signal was audible. The results of Olive and Spickenagel were largely confirmed in this experiment.