Neural network use in a non-linear vectorial interpolation technique for speaker recognition

We describe a technique for speaker identification. In this technique, the within as well as between input speech signal vector correlations are supposed to be speaker specific and are estimated using nonlinear vector interpolators. Three models were developed depending on whether the correlation between the odd-even vectors of the sequence or between the central vector and the remaining vectors or between the odd-even components of the vectors is used. For each model an interpolation function is defined which is implemented using multilayer feed-forward neural networks optimized by genetic algorithms. We have evaluated the performance of the three models (nonlinear interpolator models) using continuous speech corpus. In a test with 72 speakers, using the training data and 12 LPCC-derived cepstral coefficients as parametric vectors, all three models showed a great capability of representing the temporal correlation between sequences of speech pronounced by the same speaker after a training step. The odd-even vector model gave the best global recognition rate