Search

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.

9 Aug 2005: Particular importance must be given to the effect of words en-countered in the test data which have not been seen in the trainingdata. ... We also used 3 hours of test data from the NIST 1998 Hub-4BN benchmark tests.

19 May 2005: Among all the adaptivelytrained systems, the SATSPAM outperformed all the other sys-tems on both test sets. ... 3. Three BN test sets were used, each of them contains six 30 minutesbroadcast news shows.

26 Sep 2005: Here. P (šn|xt, M̌) P (šn|yt, M̌) (10)the model M̌ is now matched to the test condition rather. ... All results are quotedaverage of three of the four available test sets, Feb’89,89 and Feb’91.

26 Mar 2005: SID stage - Adding Eval Data to UBM. • New UBMs, BE and BMAPE, were created using the whole test data set.• System gender labels were used. ... no real cheating but against eval rules).• Using just the target show (rather than whole test set) did

9 Aug 2005: At each update a t-test isperformed, determining to a desired level of confidence whetherthe measured mean is lower than would be expected were the can-didates to occur independently. ... The following table gives a brief description ofeach corpus, where

12 Apr 2005: Acoustic Model Test Data:– eval03 6 hours (3 hours Switchboard2 Phase 5, 3 hours Fisher)– dev04 3 hours Fisher data. • ... Experimental Setup. • Acoustic model data sets:– Training data: 76 hours h5etrain03sub & 296 hours h5etrain03– Test data

9 Aug 2005: The test-set wassubjected to the same preprocessing steps as the training corpus. ... The OOV rate for the describedvocabulary is 0.62% on the test-set.

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: Test-ing was performed on 164 images containing 193 cars,and164 background images. ... Our technique relies on edge features andhigher-resolution training and test images would certainlyimprove results.