Reverse Engineering Vector Quantizers For Repartitioned Signal Spaces

In this paper we describe a method to reverse engineer vector quantizers (VQs) for the purposes of optimally repartitioning and redesigning them. Our interest in this topic has been motivated by our desire to develop a scalable version of the transform domain weighted interleaved vector quantization (Twin VQ) audio coding algorithm - part of the MPEG 4 standard. Repartitioning and redesign is accomplished by synthesizing a training set from the original VQ and then using it to train the new VQs for the different repartitioned spaces. We also present mathematical justification and experiments to support the technique used.