Nonstationary lattice quantization by a self-organizing neural network

Abstract The use of a self-organizing neural network as a vector quantizer in the case of a nonstationary lattice is considered. The nonstationarity is handled by expanding the time-dependent parameters of the lattice into a suitable base. Several experimental results are presented concerning the behaviour of the neural network in vectorially quantizing the parameters of the nonstationary vocal tract modeling of speech. The results show how the neural network is able to reconstruct the spectral model also in the case of speech segments not previously used in the training phase, thus evidencing the inherent ability of the neural network to generalize.