Search

9 Dec 2006: For exam-ple, speaker models can be built for people who are likely to be inthe broadcast (such as prominent politicians or main news anchorsand reporters). ... However, this method relies on previously seen exam-ples of the test speakers, which is not

9 Dec 2006: For exam-ple, speaker models can be built for people who are likely to be inthe broadcast (such as prominent politicians or main news anchorsand reporters). ... However, this method relies on previously seen exam-ples of the test speakers, which is not

22 Nov 2006: For exam-ple, the combination of the S2 and S3 GAUSS systems gave a 0.3%and 0.2% absolute improvements oneval03m and eval04 re-spectively.

22 Nov 2006: It does this by maximising the geometric margin – the dis-tance between the decision boundary and the closest training exam-ples – between classes.

23 Feb 2006: However discriminative training1 has also been exam-ined within this framework [3, 4].

22 Nov 2006: The output of thismodel is a one-dimensional log-likelihood. To capture thedifferences in generative process between different exam-ples, each example is then mapped to the log-likelihoodgradient-space. ... Consider, for exam-ple, the second derivatives

23 Feb 2006: For exam-ple, the combination of the S2 and S3 GAUSS systems gave a 0.3%and 0.2% absolute improvements oneval03m and eval04 re-spectively.

14 Sep 2006: The results we show later, and the exam-ples in figure 3 confirm that there is indeed a practical advantage, but it is instructive toconsider on an intuitive level what the

19 Mar 2006: W1 was found to be very large). Exam-ples are shown in Fig.

21 Nov 2006: This is best illustrated by exam-ining maximum margin training of a univariate Gaus-sian class-conditional distribution.