Search

20 Feb 2018: Trial # users average # calls median # callsAMT 140 6.5 2Cambridge 17 24.4 20. ... vol. 24, no. 2, pp.150–174, 2010. [7] Amazon, “Amazon Mechanical Turk,” 2011.

3 Jul 2018: 24,000 fishermen a year. Mostly in storms. And not every country keeps accurate records.

4 Nov 2018: 23] L. Breiman. Bagging predictors. Machine learning,24(2):123–140, 1996. [24] O. Siohan, B. ... IEEE/ACM Transactions on Audio, Speech,and Language Processing, 24(8):1438–1449, 2016. Introduction. Graphemic English systems.

3 Jul 2018: training set of Aurora4. Model A B C D AVGDNN/clntr 2.71 43.00 24.06 58.66 45.48VDCNN-max-4FC/clntr 2.32 35.99 21.20 ... 24, no. 12, pp. 2263-2276, Dec. 2016. Contact: j.m.rownicka@sms.ed.ac.uk.

20 Feb 2018: error rate of 33.2 %, and for the first and secondorder derivatives the error rates of the classifiers are 33.1 %and 24.2 %, respectively. ... 24.8 42.4par 22.7 13.1 21.7 32.4 32.7 25.2 27.4 45.1.

15 Jun 2018: Sig-nificant WER improvements were observed after interpolatingwith the n-gram LM for n-best rescoring – a common practicefor speech recognition [8, 24, 25]. ... 24] S. Kombrink, T. Mikolov, M. Karafiát, and L. Burget, “Recurrentneural network

20 Feb 2018: Per-turn. WER. Per-dialog WER. None 2 6 24 83 % 0 % 0 % Low 4 12 48 83 % 32 % 28 % Med 4 12 48 77 % 46 % 41 % Hi 2 6 24 ... Dataset. Metrics (task & user sat). R2 Significant predictors. ALL User-S 52 % 1.03 Task ALL User-C 60 % 5.29 Task – 1.54

15 Jun 2018: LM rescoredev eval. Vit CN Vit CNng4 - 23.8 23.5 24.2 23.9. ... LM #succ words dev evalng4 23.8 24.2. uni-rnn - 21.7 22.1.

20 Feb 2018: Computer Speech and Language,24(4):562–588. O Tsimhoni, D Smith, and P Green. ... Computer Speech andLanguage, 24(2):150–174.

15 Jun 2018: The stage onesystem used an HTK [24] configuration that had been previ-ously employed for all Babel tasks [25, 26], multi-genre En-glish broadcast transcription [27] and many others. ... 25, no. 3, pp. 373–377, 2017. [24] S. J. Young, G.

20 Feb 2018: wwpos 24.52 11.29 18.47wspos 11.33 4.91 3.31wpof s 1.13 4.82 8.82wppof s 24.27 6.49 10.54wonset 15.08 0.33 ... 47.3% 52.7%. 75.3% 24.7%. Figure 3: Categorical quality ratings for spectral conversion duration conversion HMM-based contour generation.

15 Jun 2018: In [24] phonetic pronunciation features consisting ofa set of phone-pair distances were proposed for vowels and ap-plied to read speech. ... 8, no. 4, pp. 369–394, 1994. [24] N. Minematsu, S. Asakawa, and K.

20 Feb 2018: The Knowledge Engineering Review, Vol. 00:0, 1–24. c 2006, Cambridge University PressDOI: 10.1017/S000000000000000 Printed in the United Kingdom.

15 Jun 2018: Word levelconfidence scores are returned from the Kaldi [24] decoderwhich are frame weighted and undergo a piece-wise mappingfor use in error detection. ... 3660–3664. [24] D. Povey et al., “The Kaldi Speech Recognition Toolkit,” in Proc.of the

20 Feb 2018: 3.3. The agenda-based simulated user. The agenda-based user simulator [24, 25] factorises the user stateinto an agenda and a goal. ... 23] TopTable, “TopTable,” 2012, https://www.toptable.com. [24] J Schatzmann, Statistical User and Error Modelling

15 Jun 2018: 24,no. 8, pp. 1438–1449, 2016.

20 Feb 2018: In control tests by humanusers, the success rate of the system was 24.5% higher thanthe baseline Lets Go! ... Com-pared to the BASELINE system, the BUDSLETSGO systemimproves the dialogue success rate by 24.5% and the worderror rate by 9.7%.

20 Feb 2018: 24, no. 4, pp. 562–588, 2010. [13] M Gašić, C Breslin, M Henderson, D Kim, M Szummer,B Thomson, P Tsiakoulis, and S Young, “POMDP-based dia-logue manager adaptation

20 Feb 2018: SFRName Reward Success #Turnsbest prior 8.66 0.35 85.40 2.19 8.32 0.20adapted 9.62 0.30 89.60 1.90 8.24 0.19. ... The systemwas deployed in a telephone-based set-up, with subjects recruited via Ama-zon MTurk and a recurrent neural network model was used

20 Feb 2018: The static feature set comprised 24 Mel-Cepstral coefficients,logarithm of F0 and aperiodic energy components in five frequency. ... 12 sentences werethen randomly selected to make up a testset for each listener, leadingto 24 wave files pairs (12 for