This research project proposes to compare and evaluate recent signal processing techniques for speaker identification. For both steps of the process, i.e., (i) features extraction and (ii) classification, new tools will be used. During the former step, a new tool for similarity analysis, called Discrete Shapelet Transform, will be adopted in addition to the traditional ones, i.e., Discrete Wavelet Transform, Cepstrum, MFCCs, and Fractal Dimension. For the latter step, current systems based on Support Vector Machines, Multilayer Perceptrons, Radial Basis Functions, and Bayesian Classifiers will be compared to two new classifiers: (i) Paraconsistent Complex Neural Networks, which have not yet been used for this purpose, and (ii) Optimum Path Forest, which is completely new for this kind of application.
News published in Agência FAPESP Newsletter about the scholarship: