Search

22 Jun 2007: Table 3 shows the P1-P2 perfor-mance3 of these baseline models on the BC and BN tests sets. ... Table 6. Unadapted %CER onbcmdev05. The models were then evaluated on the BC test set,bcmdev05.

20 Jan 2007: All the test sets comprise ofshows where a show is a broadcast of a TV or radio program. ... 3.4. Results. The tests were carried out on a BC test set (dev05bcm: 3 hours,30k words, 5 shows) and two BN test sets (eval04: 1.3 hours, 11kwords, 3

15 Aug 2007: Thispaper shows that multistyle models benefit from VTS compensationor Joint uncertainty decoding by reducing the mismatch betweentraining and test. ... Figures are reported against thedev03 test set encompassing3 hours of shows from six different news

4 Apr 2007: For all systems, the test setsupervision was generated using the MPE-SI models. ... In the following experiments, pairwise significance test wasdone using NIST provided software - , where signifi-cance tests were using a standard approach [25].

22 Jun 2007: For theBN test set, cross-adaptation in either direction gave gains of about0.5% absolute over the best individual system. ... The first half of this test set was used for tuning the translation.

10 Oct 2007: Table 3 shows the P1-P2 perfor-mance3 of these baseline models on the BC and BN tests sets. ... Table 6. Unadapted %CER onbcmdev05. The models were then evaluated on the BC test set,bcmdev05.

15 Aug 2007: Each ofthe 16 conditions also has a test set of a 1001 sentences with 52male and 52 female speakers. ... Exact and Approximate matrix inversion for PCMLLR onAURORA 2 test set A, averaged across N1-N4, WER(%).

22 Feb 2007: Classification error on the TIMIT core test set. •

10 Oct 2007: 3.3. Test datasets. The test sets used to evaluate the systems are bnmdev06 and bcmdev05.bnmdev06 comprises 3.6 hours of data taken from a range of BNsources. ... It includes some of the standard existing test sets describedin [10], dev04f, eval03m and

5 Oct 2007: The same five test sequences were also reconstructed by automatically resolving the multipledistance ambiguity. ... As a final test, we recorded two freehand sequences, again on the same speckle phantom usedfor the calibration.

22 Feb 2007: Require good generalisation– Can reduce training set error rate: need to reduce test-set errors!– Need to keep gains with large numbers of parameters– Need to increase “confusable” data for training– ... simple ASR systems - larger gains–

10 Oct 2007: For theBN test set, cross-adaptation in either direction gave gains of about0.5% absolute over the best individual system. ... The first half of this test set was used for tuning the translation.

22 Feb 2007: known phone boundaries) perfor-mance on the TIMIT core test set is shown in table II. ... of alanguage model (tests on MMI HMMs suggest that this mayyield a gain of up to 0.5% absolute).

23 Apr 2007: In iterative MLLR, a transform isestimated using the 1-Best hypothesis of the test data assupervision. ... However, in some applications, test data becomesavailable gradually. Incremental inference is often used.

10 Oct 2007: Only 0.3% improvement wasobtained on this test set. 3See http://htk.eng.cam.ac.uk/docs/cuhtk.shtml4Significance tests were carried out using the NIST Scoring Toolkit. ... These systems wereevaluated on thedev03 andeval03 test sets, each consistingof 3

21 Sep 2007: tic mismatch between training and test conditions. This report examines the application of. ... The joint distribution mayalso be thought of more generally as modeling training and test conditions jointly rather thanas clean and noisy data.

30 May 2007: known phone boundaries) perfor-mance on the TIMIT core test set is shown in table II. ... of alanguage model (tests on MMI HMMs suggest that this mayyield a gain of up to 0.5% absolute).

16 Nov 2007: data to build HMMs system with the same acoustic conditions as the test data. ... to compensate for the mismatch between training and test conditions by modifying the model.

3 May 2007: This was significantlybetter at a 95% confidence level using a pair-wise significance test. ... On both test sets the new form of model performed significantly better than standardHMMs.

16 Nov 2007: representation 103. 6.5 Synchronous stream system with confidence weighting 107. 6.6 Results using GMM features on WSJ corpus and CSRNAB hub 1 test set 108. ... diagonal transforms 127. 8.4 Experiments using constrained MLLR transforms for WSJ test