A Low-Complexity Two-Stage Adaptive Vector Quantizer.

We consider a two-stage adaptive Vector Quantizer (VQ) that performs well on a wide variety of sources at rates of approximately 2 bits/sample. The sources considered are sources with memory such as Gauss-Markov sources, speech, still images, and music. The design of the two-stage adaptive VQ is guided by high-resolution quantization theory which analyzes the role of various quantizer features (cell shape and point density) in the performance of the VQ. We compare the performance of this scheme against Jayant's adaptive quantizer with one word memory.