Search

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.

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

16 Nov 2007: 8.1.1 Training and Test Corpora 102. 8.1.2 System configurations 103. 8.2 Initial Experiments 106. ... ments. A summary of the history of NIST’s benchmark tests are given in [91].

16 Nov 2007: 89. LIST OF FIGURES xiv. 7.3 Word error rate for the 1200 utterance test data against the number of hypothe-. ... Oracle: 1.18 (test), 0.08 (train). Idiot: 63.43 (test), 52.45 (train). 95.

16 Nov 2007: the WER on test data may be viewed as the ultimate aim, or a zero risk complexity control.

16 Nov 2007: test error rate ideally. error rate. mimics theoretical. training. test. classifier complexity.

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.

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

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.

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