Search

9 Aug 2005: Training- and test-sets werecreated by splitting the material evenly acro ss these topics in the. ... A 65K vocabulary was used to build language models. Thestandard 2.1 million word set-aside dev-test text for WSJ0 was used as a test-set.

9 Aug 2005: A good approximate loglikelihood score was found empirically. To test whether the GS performance degrades below the 1.6 tailthreshold due to the state likelihoods of components within a se-lected

9 Aug 2005: Section 5. evaluates methods using the SQALEUS-English test set and Section 6.

26 Sep 2005: to test it. This was recognised early on, thus most noise robustness methods can be classified under several. ... Test Conditions. Figure 3.1: Methods of reducing the acoustic mismatch. 3.1 Inherently Robust Front-Ends.

9 Aug 2005: c) aspectratio = 1:1 (201 points). We do a further test by varying the aspect ratio of thepreattentive filter.

8 Aug 2005: Thisis the underlying principle of the Likelihood Ratio Test (LRT). It could be easily shownthat using likelihood ratios minimizes the Bayesian error in two-class decision problems[20].

9 Aug 2005: citation words, number sequences and prose passageswere used, with one quarter of the data ( 1500 phonemes) setaside as a test set. ... 6. RESULTSResults for both phoneme and word recognition over 50 test utter-ances with.

9 Aug 2005: hood estimate, then an OpenGL implementation is possible.The model is repeatedly rendered using the 0 -test, with anincreasing probability threshold.

8 Aug 2005: We performed 25 recognition tests, using each database for training and testing it against all the others.For each person in a database we collected a data set consisting of

9 Aug 2005: Test-set perplexity. Training-set perplexity. Test-set, count thresholdprune. 1e-4. 5e-5. 2.5e-5. 1e-5.