Sequence kernels for Speaker Verification using Support Vector Machines
par/by Jérôme Louradour
IRIT,
Université Paul Sabatier
Automatic Text-Independent Speaker Verification consists in determining whether a speech utterance was pronounced or not by a target speaker, without any constraint on the speech content.
We will discuss on the application of Support Vector Machines (SVM) to this binary classification task of variable-length sequences.
The approaches of interest exploit kernels that can handle sequences, and not items within sequences: both theoretical and practical reasons justify the effort of devising such sequence kernels. More precisely in our case, the sequences are sets of acoustic vectors.
Several families of kernels between sets of vectors will be reviewed. Their applicability for a speaker verification task will be discussed. Empirical results will be exhibited on a large benchmark test set.