FIT Božetěchova 2, seminární místnost UPGM L220 20.2.2009

Seminář je organizován Ústavem počítačové grafiky a multimédií Fakulty informačních technologií Vysokého učení technického v Brně. Je zaměřen především na počítačovou grafiku, zpracování obrazu, zvuku, řeči a textu. Seminář se koná na  VUT FIT Brno, Božetechova 2, L220, obvykle v pátek v 10:00.


In the standard approach of speaker verification, each speaker in the training set is represented by their own model.
IN the testing phase, given the test feature vector and the speaker model,
the task is to decide whether the features were generated by the model depending on the output score.
In this work,
only two models are trained.
Given two speech segments, encode them into one feature vector which is denoted as a
trial feature vector. It results in having two sets of trial features: target trial features and nontarget trial features.
Distribution of each set is then represented by one model. Given an input test trail feature vector, the task is to decide
whether it was generated by the target trial model or the nontarget trial model.