Robust vector quantizer design using self-organizing neural networks

In this paper we propose a new method to design vector quantizers for noisy channels. Self-organizing neural networks are known for their efficiency in voice and image data compression; we use self-organizing algorithm to create a topological similarity between the input space and the index space. This similarity reduces the effect of channel noise because any single bit error in a transmitted index will be translated to a close codevector in the input space which yields relatively small distortion. For an 8-bit vector quantizer, the proposed system resulted in 4.59 dB spectral distortion in a highly noisy channel while a simple LEG, LEG with splitting and 2-D self-organizing map (SOM) resulted in 5.96, 5.46 and 5.02 dB of distortion, respectively. (C) 2000 Published by Elsevier Science B.V. All rights reserved.