A new robust hybrid speech recognition algorithm based on FVQ/HMM and neural nets classification

In this paper we propose a new robust hybrid isolated word speech recognition system which combines the improved quantization accuracy of Fuzzy Vector Quantization (FVQ) with the strength of HMM in modelling stochastic sequences, and the non-linear classification capability of MLP neural networks. The proposed FVQ/HMM/WLP system approach investigated the relative contributions of codebook-dependent Fuzzy distortion information and probability-dependent matrixmum likelihood probability information The scheme is effectively a powerful combination of the FVQ/HMM and FVQ/MLP systems. Computer simulation results clearly indicate the superiority of the FYQ/HMM/MLP approach over that obtained from FVQ/HMM or FVQ/MLP.