Search

10 Jan 2005: The transcription is either known in training orobtained by an initial, non-VTLN, decoding pass for test data.If and are the original and transformed feature vectorsrespectively then the log-likelihood ... The systems were eval-uated on two 3 hour test

9 Aug 2005: We also used 3 hours of test data from the NIST 1998 Hub-4BN benchmark tests. ... Most of these words are not capitalised,if they are used in the middle of sentences.

31 Jul 2000: Pronunciation modeling by sharing Gaussians 141. TABLE I. WER degradation with speaking style on theMULTI-REG test set. ... Steps (6), (3), (4) and (5) arethen carried out in that order to estimate and test a matched pronunciation model.

9 Aug 2005: Steps 6, 3, 4 and 5 are then carried out to estimateand test a pronunciation model. ... in Step 5 (Test) PER WER PER WERNone (Dictionary) 49.1% 49.1% 49.5% 58.9%Tree Pron.

8 Aug 2005: These meth-ods have achieved very good accuracy on a small number ofcontrolled test sets. ... We therefore useDKL(p(0)||p(i)) as a “distance measure” between trainingand test sets.

12 May 2003: Top and middle row: model pro-jected on images from camera 1 and 2, respectively. ... 4. Experimental Results. Real data experiments were designed to test the proposedtracking algorithm.

17 Feb 2006: In AFR tests, such methods are usuallyoutperformed by methods from the third class: view-based techniques e.g. ... KLD) [38]. In all tests, both training data for each person in the gallery, as well as test data,consisted of only a single sequence.

25 May 2007: Wetherefore also test a 2D colour map in which strain is indicated by intensity, and a colour (red inour example) is introduced to indicate precision. ... Middle row: corresponding pseudo-strain images without spa-tial averaging. Bottom row:

12 May 2003: Point correspondences betweentwo shapesare found by minimizing the point matchingcosts,which is the test statistic for histograms. ... 7. Figure 5: Results of hand localization. Left column: handlocalization usingshapecontext informationonly

3 Apr 2006: and used for training wereclean, single-instance images, and so our performance overthe 150 remaining test images would be therefore expectedto be slightly lower than for the whole database. ... Left: input image. Middle: inferred parts la-bellings.