Feature parameter curve method for high performance NN-based speech recognition

This paper proposes a novel high-performance NN-based isolated word speech recognition method named feature parameter curve method (FPCM), which contains three sub-methods, i.e., the feature frame parameter method, the broken line fitting method and the representing-point sequence (RPS) method. Using feature parameter curve to represent the contour of LPC-based cepstral coefficient sequence, our new methods solve the time alignment problem, greatly lighten the burden on neural network and thus reducing the complexity of the neural network. Compared with other NN-based methods, our methods have many strong points such as much faster training and recognition, simpler structure of neural nets, and is more robust. A recognition system of all Mandarin syllables has been established based on this method and the comparative experiments confirmed the good characteristics stated above. The idea of reconstructing original speech signals is also first shown in speech recognition.