Topological Representation of Speech for Speaker Recognition

Abstract: This research is in the field of speaker recognition and is based on the fact that behind each supervised or unsupervised learning algorithm, there is an implicit assumption about the structural nature of the data. Our proposal is to find new speaker information using models of his speech. In order to characterize the speaker by his speech, a mapping function can be obtained from a set of speakers models, then projecting any new speaker in a new space which increases the variability between classes and decreases the variability intra classes, for its classification. We will use topological space as a support to use inner geometric structure created by the projection of the acoustic models over a new spaces that reflects the nature of the speech acoustic classes of each speaker