Search

He has authored and co-authored more than 200 papers.

27 Jul 2020: Abstract for rosti_csl04. Computer Speech and Language, 18(2):181-200, 2004. FACTOR ANALYSED HIDDEN MARKOV MODELS FOR SPEECH RECOGNITION.

23 Feb 2002: Proc. TREC-7, pp. 191-200 (Gaithersburg, Maryland, 1998). 6. S.E. Johnson [Reducing Word Error Rates of Found Speech: XPERT Tool for Transcription Analysis.

17 Nov 2017: Eval{1,2,3} sets (about 13 hours) contain spontaneous speech from 200 speakers with Gujarati, LA Spanish and mixed L1s, respectively. ... 5. Parse Tree. 7. Conclusion. Trn3 set contains trn1 and 200 hours selected from the unsupervised set.

27 Jul 2020: Semi-automatic image processing facilitates alignment of the beam and wires to within a tolerance of around 200 microns, despite the considerable beam thickness.

17 Nov 2017: c) (d). 0 50 100 150Frame Index. 0. 100. 200. 300. ... 100. 200. 300. 400. Frequency. (H. z). k 0 = 3 k 0 = 5 k 0 = 7.

28 Apr 2005: Error Analysis. c.200 DEL errors analysed by hand:. • 40.7 %: before asyndetic clause boundaryEx: they destroyed all the national monuments () he destroyed a large area. • ... c.200 INS errors analysed by hand:. • 26.2 %: before potential

17 Nov 2017: 403. 404. 50. 100. 150. 200. 250. 300. 350. 3762 7. ... I Cumulative state data usage statistics for 3-hour configuration (VLLP). 100 200 300 400 5000.

30 Jan 2000: Images 1 and 5200 0 200. 400. 200. 0. 200. Images 2 and 3. ... 200. Images 4 and 5. Figure8: Imagesof the ellipsesandepipolarlines after convergence.Eachfigure showsthe overlappingof images and , for -GP& 555 # and ¤@@P 555.

10 Dec 2003: 0. 0.5. 1. 1.5. 2. 0 200 400 600 800 1000. % ... 0. 0.2. 0.4. 0.6. 0.8. 1. 1.2. 1.4. 1.6. 0 200 400 600 800 1000.

17 Nov 2017: BBC Monitoring 200 journalists monitoring global audio, video, text news. Deutsche Welle Multilingual content generation and management.

8 Mar 2000: 100 samples were used in the Monte Carlo simulations. Table 1 shows the average one-step-ahead prediction errors obtained for 10 runs, of 200 time steps each, on a Silicon

23 Mar 2004: Key Issues: 2004 Data. • New data planned for collection in 2004 includes– more Fisher English CTS data– 200 hours of Levantine Arabic CTS data– 200 hours of Mandarin CTS data– English

3 Jul 2018: Results. # of Clusters 50 100 150 170 200 TextTiling [Hearst 1997] 0.484 DNN-HMM [Yu et.

14 Aug 2000: BRF is run on the parallel collection adding the best 200 new terms taken from the top 10 documents to the original document with a term frequency equal to 1.

25 May 2004: Gaussianisation. • Transform any distribution to standard Gaussian N(0, I). 20 15 10 5 0 5 10 15 200. ... 0.01. 0.02. 0.03. 0.04. 0.05. 0.06. 0.07. 0.08. 0.09. 0.1. 20 15 10 5 0 5 10 15 200.

20 Nov 2017: Linear Conjugate Gradient. Experimental Setup. I Experiments used ASRU 2015 MGB challenge data.I Training set : 200 hr MGB1 training dataset.I Validation : the official MGB1 dev.sub set (5.5

23 Feb 2002: s1. -. -. 12. 26. 59.11. 57.14. 54.04. 50.65. s1. 200. ... 10. -. -. 50.76. 49.42. 52.91. 51.67. s1. 200. 10. 8.

19 Jan 2018: rate thanthe 200 frames-per-second typical in many sys-tems.

7 Oct 2009: Frame accuracies are given foreachof the 200 hours subset MLPs in Table 1 and all subset networksgive similar results. ... 1The best performing 200 hours subsets MLPs.2The GALEeval07 non-sequestered testset version is used.

18 Jan 2011: a) (b). (c) PHONG2D (d, t, µ, n = 200) (d) PHONG2D (d, t, µ, n). ... a) (b). (c) PHONG2D (d, t, µ, n = 200). (d) PHONG2D (d, t, µ, n).

23 Dec 2017: To builda speaker dependent lipreading system, we can use, for example, 800 utterances as trainingsamples and 200 as test samples. ... and we wouldtest on the 200 samples of speaker Y and vice versa.

10 Apr 2000: 100 samples were used in theMonte Carlo simulations. Table 1 shows the average one-step-ahead prediction er-rors obtained for 10 runs, of 200 time steps each, on aSilicon Graphics ... 10. 12. 34. 0. 50. 100. 150. 2000. 100. 200. 300.

7 Oct 2009: The 260K word-list. Training data Test Acc. (%)200 hours 63.551350 hours 65.16. ... Table 1. Frame accuracies of the 200 hours and the 1350 hours MLP.

8 Mar 2000: 0. 50. 100. 150. 200. 250. 300. 350. 0 50 100 150 200. ... 0. 200. 400. 600. 800. 1000. 1200. 0 50 100 150 200 250 300 350.

17 Jan 2003: 100. 200. 300. 400. 500. 600. 700. 800. Figure 1: Distribution of #Gauss among states. • ... LogDet> -0.0051<BlockInfo> 1 3<Block> 1. <XForm> 3 31.01 0.30 0.200.10 0.97 0.200.10 0.20 1.10.

30 Mar 2000: Postscript Version : Proc. TREC-7, NIST SP 500-242, pp. 191-200 (July 1999). ... Misrecognising every word will in practise give a TER of below 200% as the word ordering is unimportant, so some recognition errors will cancel out.

23 Mar 2004: 0.8. 1. 1.2. 1.4. 1.6. 0 200 400 600 800 1000.

3 Aug 2010: 10. 020. 30. 40. 50. 60. 70. 100 200 300 400 500 600. ... 11. 020. 30. 40. 50. 60. 70. 100 200 300 400 500 600.

9 Aug 2005: Fre. quen. cy (. kHz). 0 200 400 600 8000. 2. ... Time (ms). Fre. quen. cy (. kHz). 0 200 400 600 8000.

9 Aug 2005: 100. 150. 200. 250. 1. 50 100 150 200 250 300 350. ... 50. 100. 150. 200. 250. 0 50 100 150 200 250 300 350.

22 May 2006: SENSORLESS RECONSTRUCTIONOF UNCONSTRAINED FREEHAND. 3D ULTRASOUND DATA. R. J. Housden, A. H. Gee,G. M. Treece and R. W. Prager. CUED/F-INFENG/TR 553. May 2006. University of CambridgeDepartment of Engineering. Trumpington StreetCambridge CB2 1PZ.

14 Sep 2006: The question then is to what extent adding marginal samples. 0 50 100 150 200 250 3002.

23 Nov 2003: 0 50 100 150 200 250 300. 0. 10. 20. 30.

22 Sep 2005: 0 200 400 600 800 1000 12000. 0.5. 1. 1.5. 2.

20 Apr 2005: In some of the experimental sequences,the displacement expert loses track. On these occasions, theRMS error becomes meaningless (being wrong by 100 pixelsis no better or worse than 200 pixels). ... These results were taken from tracking a 200 frame

21 Sep 2006: Furthermore, since the size of the ROIs may vary from one videoto the other, the final MGO images are resized to the corresponding refined images withstandard size (which is 200 200 ... 3.2 Dimension Reduction by PCA. In order to reduce the number of

19 Feb 2007: 10. 250 200 150 100 50 0 501.5. 1.6. 1.7. 1.8.

6 Mar 2008: 0 50 100 150 200 2500. 0.2. 0.4. 0.6. 0.8. 1. ... Although the artefact’s appearance. 11. (a1) (a2) (b1) (b2). 100 150 200 250 300 350.

9 Aug 2005: 0 1000 2000 3000 4000 5000 6000 7000 8000300. 200. 100.

9 Aug 2005: A4 68@BH0. LP(10) Residual. -500. 0. 500. 100 200. LP(10) Residual Spectrum. ... 2000. 100 200. NLP(10-2) Residual. -500. 0. 500. 100 200. Original Spectrum.

5 Dec 2003: 50 100 150 200 250. 160. 180. 200. 220. 240. 260. ... 200. 220. 240. 260. 280. 300. 320. 340. 100 101 102 103 104 105 106 107 108 109.

15 Feb 2006: 9. (a)0 100 200 300 400 500. 0. 2. 4. 6.

9 Aug 2005: True. Pos. itive. Rat. e. ROC curve for diphone bee for female speaker.0 50 100 150 200 250 300.

20 Aug 2009: Table 2: RMSE(mm) for the in vitro experiments. A total of 200 trials were performed on 5 datasets. ... For each similarity measure, 5 simulations were performedper dataset (see Table 2) giving a total of 200 simulations over four optimisation algorithms

10 Oct 2007: IEEE TRANS. ON SAP, VOL? , NO? ,? 200? 1. Automatic Model Complexity Control Using. ... 200? 7. as the average phone correctness of all possible word sequences {W̃},.

18 Dec 2003: TheCallHome English (CHE) corpus consists of a total of 200 conversations between family membersand close friends, no restriction was placed on the conversation topics. ... Criterion Training set τ I train eval98ML h5train00sub — 47.8 46.5MMI

8 Mar 2000: Note, not producing any output gives a TER of 100%whereas misrecognising every word as on OOV word produces a TERof 200%, due to each substitution error counting as both an ... Misrecognising every word will in practise givea TER of below 200% as the

10 Apr 2000: C,"understand","understand",413.300413.880:C,"why","why",413.880414.080:C,"it","it",414.080414.200:C,"makes","makes",414.200414.420:C,"a","a",414.420414.490:C,"

9 Aug 2005: log. spec. trum. ÐA[_nZGaIlhji[WcVfjap]VQds]£9apVQ! CxQ$agZ rappjiZmW1xnLÂf]bU$xnxnZGap]Mq@xn[haLº. 0 50 100 150 200 250 300 350 400 450 5000.