Search

14 Jun 2004: Test Data. to be completed. For testing with some decoders (e.g. ... ge204@watch$ /home/solveb/hub5/htkbin/bin.i686/HRConvert.090302 /home/widowb/ears/bn-E/exp/eval98-test/hmms/hrconvert.cfg -a hmm164/MMF -b xwrd.clustered.mlist -c hmm164/RMF |

14 Jun 2004: This was to simulate conditions in the RT-03s evaluation, where there was a temporal gap between the test-audio and the training shows. ... To help minimise the cost of computing the BIC statistics at every point, a faster Hotelling's T test is first

23 Mar 2004: Woodland: STT Overview. Measuring STT Progress. Current Test Sets : Traditional NIST speech recognition evaluation approach. • ... Progress Test Set : Fixed test set to eliminate test-set variability for trackingprogress towards program goals. •

13 Jul 2004: NIST,. Benchmark Tests : Rich Transcription (RT).

25 May 2004: Test setseval03 6h set from Fisher and Swb2-5 data, 72 conversationsdev04 3h set from Fisher, 36 conversations. ... WER on eval03, fsh2004sub models, unadapted, 2003 trigram. Test with adaptation in P1-P2 system:.

23 Dec 2004: For this work both the dev04 and eval03 test sets were usedfor development. ... The new LM gave 1.1%absolute reduction in CER on the dev04 test set.

27 May 2004: During training and test adaptation, a global in-terpolation weight was estimated and separate speech and silencetransforms were used for CMLLR. ... Using ST for test adaptation again yieldedabout the same as a GD system with CMLLR.

9 Dec 2004: Corpus consists of– 48 training examples per vowel (total of 528 examples)– 42 test examples per vowel (total of 462 examples). • ... φcn(O; λ) =. F(ω1) F(ω2). φ(O; λ)1. . • Incorporating in score-space requires consistency

29 Apr 2004: Test on different systems, architectures and data sets. • Allow more than 2 inputs. ... Use dual-phone recogniser to remove long silences. • Use LIMSI-style segmentation/clustering algorithm [ NB - parameters havebeen chosen on the test data].

18 Jun 2004: During training and test adaptation, aglobal interpolation weight was estimated and separate speech andsilence transforms were used for CMLLR. ... Using ST for test adaptation again yieldedabout the same as a GD system with CMLLR.