Combining neural network classification with fuzzy vector quantization and hidden Markov models for robust isolated word speech recognition

This paper proposes a new robust hybrid isolated word speech recognition system which is based on the improved quantization accuracy of FVQ, the strength of HMM in modelling stochastic sequences, and the nonlinear classification capability of MLP neural networks. Thus the proposed FVQ/HMM/MLP approach combines effectively the relative contributions of the codebook-dependent fuzzy distortion measures with model-dependent maximum likelihood probability information. Computer simulation results clearly indicate the superiority in recognition accuracy performance of the FVQ/HMM/MLP approach when compared to that obtained from FVQ/HMM or FVQ/MLP schemes.