An algorithm for designing optimum quantizers subject to a multiclass distortion criterion

This paper describes an extension of the Lloyd-Max optimum quantizer design algorithm to a multi-class mean-square-error optimally criterion. As part of this study, a comparison of uniform versus non-uniform gathering of bins in histograms used to approximate probability density studies is presented. In the associated experimental study, the new quantizer design procedure is applied to a forward adaptive ADPCM coder, and the resulting convergence, speed of computation, and final error statistics are examined. Listening tests indicate that the multi-class design procedure produces a clearly perceivable, though modest, improvement in the speech quality.