Using neural networks for vector quantization in low rate speech coders

The problem of reducing the complexity of the codebook search in low-rate speech coders is addressed. Emphasis is placed on vector quantization of the short-term parameters (spectral parameters), where the increasing demand for higher performance necessitates codebook sizes of approximately 2/sup 20/. As full search is impractical, a novel path search algorithm is proposed. it is based on a multidimensional version of Kohonen's self-organizing feature map, using the ordering aspects of the map. A comparison with the full-search LBG algorithm shows a substantial reduction in search complexity with only a minor degradation in speech quality. Furthermore, the speech quality is better than that obtained with split-LBG.<<ETX>>