The mean squared error criterion: Its effect on the performance of speech coders

Many speech coder designs use analysis-by-synthesis techniques in which a suitable output is chosen on the basis of mean squared error between input and output waveforms. A method is described for modeling the behavior of such coders, in an effort to discover what implications the use of this criterion has for the characteristics of the coders which use it. The model is used to show that as bit rate decreases, the mean-squared-error criterion leads to a progressively greater distortion of the short-term spectrum of the coded speech. The distortion contains additive and multiplicative components, with the latter dominating at low bit rates. The results underline the extent to which the criterion can shape the behavior of a coder and limit its potential performance