Quantização vetorial: aplicações a um vocoder LPC.

Abstract

The use of Vector Quantization in voice coding has been of great interest to researchers in the field, due to the possibility of coding voice signals at rates as low as 800 bits / s. This work deals with the basic principles of Vector Quantization, with emphasis and applications focused on a "VOCODER LPC". A brief description of the "vocoder" used in the simulations and three common distortion measures in vector quantization (quadratic error distortion and two versions of Itakura-Saito distortion) are presented. Also described are two reproduction alphabet design algorithms, the LBG algorithm and the "Product Code" algorithm. With these algorithms and the three distortion measures, five vector quantizers are implemented. Some methods of generating initial alphabets are also discussed, and a simple method for reducing empty cells during generation of reproduction alphabets is proposed. Objective and subjective evaluations of the implemented quantizers are performed and the results of these evaluations are presented, and for the subjective evaluations, a formal test based on the average opinion score was used.