Multilayered and columnar competitive networks for spoken word recognition

We have presented a multilayered and columnar competitive network involving competitive associative nets (CANs) and adaptive vector quantization nets (AVQNs) for spoken word recognition. Although the network has shown good performance in recognition rate, it requires a relatively large calculation time owing to the CANs. So, here, we present a new network replacing the CANs by a conventional feature extractor and an additional AVQNs, where as the feature extractor we use one of the three conventional methods: RLS (recursive least squares) for extracting LPCs, LSL (least squares lattice) for PARCOR coefficients, and normalized LSL for normalized PARCOR coefficients. As a result of experiments, the normalized LSL shows almost the same performance as the original network in recognition rate while reducing the calculation time.