SOFFA(Self-Organizing Feature with Fuzzy Association) neural network model is proposed and applied to isolated word recognition. The SOFFA neural network consists of local feature extraction, fuzzy abstraction, and MLP stages. Local feature extraction stage is trained by either competative learning or error back propagation. Two speech databases with 75 and 50 words, respectively, are used. The recognition accuracies of 91.5% and 97.5% are obtained for speaker independent test words. The performance is superior to nearest-neighbor classifier and similar to HMM(Hidden Markov Model) and MLP. The SOFFA neural network is easier in hardware implementations.
