Search

5 Oct 2000: Lattices were generated on the training set using a bigramLM. The bigram 1-best hypotheses had a 24.6% word error rate(WER) and a Lattice WER (LWER) of 6.2%.

10 May 2000: ID BASE BRF UBRF3k 22.2 24.4 33.37k 33.8 37.5 44.3. 14k 41.4 47.6 51.827k 43.0 49.7 53.855k 43.5

17 Oct 2000: 704/c 26.3 15.570a/b 10.3 1.970f/d 18.8 7.670w/c 24.2 15.8. Table 1: Lattice error rates for several development test speakerswith ... a fg 14.24 —HMM-2 thresh. b fg 13.81 —HMM-2 thresh. c fg 13.71 6.68.

10 Apr 2000: M S TM 82.11 17.89 0.00N 15.27 84.22 0.51S 0.56 98.24 1.20T 0.00 1.19 98.81. ... 1F3 36.4 32.7 35.3 33.9F4 28.6 25.0 25.4 24.4F5 28.6 23.8 27.1 26.5FX 58.5 55.2 56.8

8 Mar 2000: Y 14.2 8.0 15.4 20.3 16.5 14.0 16.6 24.6. ... ROVER fgintcat 4/Y1/N 13.8 7.8 15.1 20.1 15.8 13.6 16.6 24.1.

8 Mar 2000: 36.5F4 24.2 25.1 23.1 21.3F5 25.7 24.1 25.9 23.0FX 57.2 53.7 57.1 50.4. ... 24.1 29.9ROVER BN2/3/gd ic 1/4 15.8 9.4 15.2 19.5 26.9 19.4 22.1 29.1.

8 Mar 2000: Test Data Adaptation ClassesSubset None Global MultipleNM/FB 26.7 24.8 22.3BM/FB 41.2 32.8 31.3NM/RB 58.5 40.7 36.5Overall 36.4 ... HMM-2/1 trans fg 27.2 18.8 26.4 32.6 24.1 29.6 21.1 48.2HMM-2/2 trans fg 26.9 18.7 26.2

6 Nov 2000: Some discriminative training schemes, such as frame-discrimination [14, 24], try to over-generate training set con-fusions to improve generalisation. ... In [24] it was shownthat the improvements obtained by FD were at least as goodas those reported by

26 Apr 2000: Althoughvolume graphics (the storage and display of volumetric data as opposed to polygonal surfaces)has advanced considerably over the last decade [14, 24], most surfaces are still defined aspolygonal models, and ... the skeleton representation [2]and

12 May 2000: There are several techniques which can successfully generate this voxelarray from single sweeps [24]. ... Figure 1(a), forinstance, shows two sweeps from a liver examination which have been interpolated using thevoxel nearest neighbour [24] interpolation.