Nonspecific speech recognition method based on composite LVQ1 and LVQ2 network

A novel method of normalization is proposed in this paper, in which the MFCC (Mel frequency Cepstral Coefficient) and DeltaMFCC (difference mel frequency cepstral coefficient) are sampled equidistantly. For these normalized signals, a new speech recognition based on composite LVQ1(Learning Vector Quantization) network and LVQ2(Improved Learning Vector Quantization) network is presented. First, MFCC and DeltaMFCC feature extraction algorithms are introduced, then their coefficients are normalized. The recognition is first to learn coarsely by LVQ1 network and then to learn finely by LVQ2 network. Finally the simulation is given, which shows that the proposed algorithm improves the recognition rates effectively, with shorter training time in comparison with LVQ1 network used alone.