Search

24 Jul 2003: Backchannels not labelled separately in the training data.• Only the test data has reference ctm/mdtm files. - ... Bigram SULMs give largest reductions in %Err when combined with the PFM(using the current training and test data!). •

23 Sep 2003: For BN the test datais taken from radio and TV news shows (for example CNNHeadline News, ABC World News Tonight). ... The 2003 BNeval test set consists of 6 half-hour excerpts from Radio andTV broadcasts taken from February 2001.

16 Sep 2003: Procedure. Training only Test onlyBoth dictionaries. single variant. substitutions. insertion. deletionsorA B C. ... Pronunciation variants in training and test. Combining different strategies in training and test.

19 Sep 2003: Fig. 1. Test set word error rate for all possible models, with thestandard front-end 12, 16 and 24 component performance. ... there was some random variation in the performance,since we do not have an infinite held out test set.

20 Nov 2003: CTS SU-Detection System Overview. SULM_1 SULM_2. SULM_N. RTTM FILE. TEST DATA. ... Rebuild PFM and tg SULM.3. Obtain results for the dev03f test set.4.

25 Sep 2003: Table 1. The WER(%) on test set CEDFÑ _ for MMI/MPE systemsand discriminative SAT systems, after constrained MLLRadaptation. ... Furthermore, the useof MMI-based constrained DLT for building an MPE-SAT modelcould also get similar test-set performance.

23 Jun 2003: Acoustic Training Set-Up. • Acoustic/Training Test Data:– training data: 34.9 hours, 379 sides, from LDC CallHome (22.4hrs) and. ... dev02 numbers use manual segmentation, eval03 uses automaticsegmentation. • Comparison of dev02 and eval03

18 Dec 2003: Multiple training and test sets are used in the experiments in this paper. ... 11. Training only Test onlyBoth dictionaries. single variant. substitutions. insertion. deletionsorA B C.

24 Jul 2003: Baseline Acoustic Models: Results (I). • Development test setsBNeval98 two 1.5-hour data setsBNeval02 1-hour data setBNdev03 three hours of TDT-4 data from Jan ’01 transcribed by STT sites. • ... MLE baseline system– Add 3rd derivatives HLDA,

23 Nov 2003: It seems that some form of single stage regressionmight be more powerful, both for dealing with the rangeof variation of test examples, and for correctly modellingstatistical variability. ... Forefficiency, this test is made only every M frames (in

10 Dec 2003: Adaptive Training with Structured Transforms. Initial Results on 16-Component System. System Training Test EstimationAdaptation Adaptation MLE MPE. ... Adaptive Training with Structured Transforms. Initial Results on 28-Component System. System Training

25 Sep 2003: GD models for each bandwidth were generated. In test-ing, the reduced bandwidth models are used for transcribing dataclassified as narrow band. ... Performance. The results on thedev03 andeval03 test sets for each of thesestages are shown in Table8.

16 Sep 2003: 357 female sides (23.75hrs), only 142 male sides (9.43hrs). • Same test setup as for 20hr QT expts. ... MPE gain is larger than MLE gain. • Gains larger on female test data (biased training).

10 Dec 2003: Introduction. • Investigate linear transform parameter estimation for. – Adaptive training– Unsupervised test set adaptation– Supervised adaptation (enrollment). • ... DLT for Supervised Adaptation (II). Test sets iterations MLLR H-cri MPE-DLT.

25 Sep 2003: Previous research on LVCSRMMI training shows that the selection of Dj considerably affectsthe criterion convergence and test set generalization [8, 9]. ... An updated trigram language model was also usedin the following full decoding experiments, while

24 Jul 2003: Development test sets. dev01 40 sides Swbd2 (eval98), 40 sides Swbd1 (eval00), 38 sides Swbd2 cellular(for manual segments). ... CNC P4.[123]P5.[123] 24.1 17.1 20.7%WER on eval03 (current test) for all stages of 2003 system (auto-segments).

17 Jan 2003: Acoustic Training/Test Data. Train on the combined set of 1997 and 1998 data. • ... Development test sets. BNeval98 two 1.5-hour data sets. BNeval02 1-hour data set.

12 May 2003: 5 Experimental Results. Real data experiments were designed to test the proposed tracking algorithm.

10 Dec 2003: Training and Test Data Sets. • Acoustic training data. h5train03b 360h data set.– 290h LDC data (SwbI, CHE, Swb Cellular) with MSU/LDC careful. ... Test sets. eval02 5h set from SwbI, Swb2 and Swb Cellular data, 60 conversationseval03 6h set from

12 May 2003: 4. Experimental Results. Real data experiments were designed to test the proposedtracking algorithm.

23 Jun 2003: Narrowed gap between 100+ xRT and 10xRT considerably. • Infrastructure for quick-turnaround system tests (vs.

12 May 2003: Point correspondences betweentwo shapesare found by minimizing the point matchingcosts,which is the test statistic for histograms.

10 Dec 2003: test on eval03: skip 66% segments, 43% audio, 32% rescoring runtimei.e.

23 Nov 2003: Comparison of proposal strategies In order to test the performance of our proposalstrategies against each other we ran the algorithm for the real scene three times, usingeach strategy alone.

9 Oct 2003: The results show that when the library data is sufficiently close in epoch to the test data8 the advertdetection is very successful for some broadcasters. ... For the case where there is a gap between the library data and the test data the system is not

19 May 2003: This set of experiments therefore tests the assertion that aligning the sweeps based ona single reslice in the overlap region does not significantly prejudice the result compared withaligning the sweeps based