Search

2 Mar 2009: The ML-SAT test procedure is first run toobtain initial ML speaker transforms. ... There is no need to estimate speaker-specific discriminative transforms on the test data.

7 Oct 2009: The 260K word-list. Training data Test Acc. (%)200 hours 63.551350 hours 65.16. ... LINadaptation shows WERreductions over not adapting thefeatures for two of the test-sets.

28 Jul 2009: The running of tests consisting of all possible combina-tions of all trackers on all test sequences would take a pro-hibitive amount of time to complete. ... In order to test the validity of such a setup, weperformed tests using the complete tracking

2 Mar 2009: To further refine the LM, unsupervised test-setadaptation to a particular broadcast show, for example, may beused. ... 5. ConclusionUnsupervised test-time context dependent adaptation of n-grammixture models using a discriminative method was

16 Jul 2009: word-internal triphones for SVM rescoring models– test data collected in car (idle, city, highway), unknown length digits. ... SVM Rescoring on AURORA 2.0. System Test SetA B C. VTS 9.8 9.1 9.5+ SVM 7.5 7.4 8.1.

2 Mar 2009: increased adaptation power on unseen test data can compensatefor this and achieve overall improvements. ... Statistical significance tests showed that almostall additional gains with different MLLR improvements weresignificant13.

2 Mar 2009: For the unweighted case the finalscoreS(Ov, s) assigned to test utteranceOv is given by. ... Only the test datais used to adapt the UBMs for the parametric kernel.

18 Feb 2009: These model-basedapproaches modify the parameters of the acoustic models, so that they are representative of thetarget test environment. ... It is worth emphasising that the transform W(rm) is associated witha particular test acoustic condition.

28 Jul 2009: The running of tests consisting of all possible combina-tions of all trackers on all test sequences would take a pro-hibitive amount of time to complete. ... In order to test the validity of such a setup, weperformed tests using the complete tracking

28 Jul 2009: The most advanced algorithms stillcompete with each other using training and test data consisting of. ... The algorithm’slow pedestrian detection rate in our daylight test sequence wassomewhat surprising.

28 Jul 2009: The running of tests consisting of all possible combina-tions of all trackers on all test sequences would take a pro-hibitive amount of time to complete. ... In order to test the validity of such a setup, weperformed tests using the complete tracking

28 Apr 2009: Noisy CMLLR vs CMLLR Performance. • Evaluated on engine-on/highway noise condition from the Toshiba data– phone-numbers task (unknown digit length sequences)– see next section for test data configuration, here ... Noise Parameter Estimation. • In

20 Aug 2009: Exploration is the ability to test various regionsin the problem space to reach a global optimum and exploitation is the ability to search aroundthe candidate solution.

28 Jul 2009: The running of tests consisting of all possible combina-tions of all trackers on all test sequences would take a pro-hibitive amount of time to complete. ... In order to test the validity of such a setup, weperformed tests using the complete tracking

28 Jul 2009: The running of tests consisting of all possible combina-tions of all trackers on all test sequences would take a pro-hibitive amount of time to complete. ... In order to test the validity of such a setup, weperformed tests using the complete tracking

2 Mar 2009: 8] A. Martin, “The NIST year 2002 speaker recog-nition evaluation plan,” 2002, Available fromhttp://www.nist.gov/speech/tests/spk/2002/doc.

2 Mar 2009: 15, pp. 2465–2475, 2007. [10] A. Martin, “The NIST year 2002 speaker recog-nition evaluation plan,” 2002, available fromhttp://www.nist.gov/speech/tests/spk/2002/doc.

21 Jul 2009: As they are not re-estimated on the test data, they are not sensitive to the supervisionhypothesis errors or limited amount of data. ... The test-set eval03 consists of about 6 hours of data from144 speakers, taken from Swbd and Fisher corpora.

7 Oct 2009: MLP config. Training data Train Acc. Test Acc. 1350hr 69.51% 65.61%1350hr big 71.37% 67.57%200hr set #1 68.77% 63.92%200hr set #2 68.54% ... 4.2. Unadapted decoding results. In a first test series the use of MLP features is evaluated withinthe

9 Mar 2009: Local strain. 5This data set was previously used to test a parametric method for non-uniform smoothing: see Fig. ... The phantom data in Figs. 4, 5 and 6 was also used to test convergence for various methods ofsolving eq.

28 Jul 2009: The most advanced algorithms stillcompete with each other using training and test data consisting of. ... The algorithm’slow pedestrian detection rate in our daylight test sequence wassomewhat surprising.