A neural network quantizer for long term vocal tract characterization

Abstract

The performance obtained using a self-organizing neural network for the vector quantization of the reflection coefficients of a nonstationary lattice is considered. The training of the neural network is effected on a small number of speech patterns of one speaker and subsequently tested on different patterns of the same speaker. The use of a self-organizing neural network for quantizing the parameters representing a nonstationary lattice has evidenced an important property of this network when used as a quantizer, i.e., its inherent ability to generalize. When used in connection with speech, the network has been able to behave well in situations different from those considered in the training.<>