Etude de la réduction non linéaire de la dimension du signal de parole en vue de modélisations discriminatives par SVM
José Anibal Arias, Régine André-Obrecht, Jérôme Farinas, Julien Pinquier.

In this article we study some results of the non-linear dimensionality reduction of speech vectors. Spectral clustering, Kernel PCA, Isomap, Laplacian eigenmaps and Locally Linear Embedding are related non-supervised methods that help to discover important caracteristics from data such as high-density regions or low-dimensional surfaces (manifolds). This reduction of dimension is a necessary step when we whant to model speech sequences with discriminative functions such as Support Vector Machines.