Combining fuzzy vector quantization and neural network classification for robust isolated word speech recognition

The issue of robust isolated word speech recognition in cases where the input signal is corrupted by acoustic noise, is addressed with a new fuzzy vector quantization (FVQ)/neural network scheme. The proposed system combines in a simple and effective way the fuzzy classification capability of FVQ with the non-linear pattern discrimination power of the multi-layer perception (MLP) neural network. The paper thus defines the design and algorithmic operation of this system and compares its recognition performance to that of a conventional FVQ/hidden Markov model (HMM) system. Computer simulation results obtained using speech corrupted by car or white noise indicate that FVQ/MLP provides significantly better performance than FVQ/HMM