Search

14 Aug 2000: This parallel collection was larger and thought to be more reliable than the automatically transcribed test collection. ... A strategy proven to work well in SDR [2] is blind relevance feedback (BRF) on the test collection.

8 Mar 2000: This corpus spanned January to May 1997 and thus pre-dated the TREC-7 test data. ... 39.73. 44.15. Table 2: Average Precision on the TREC-7 SDR test collection with mean number of query terms in parenthesis.

8 Mar 2000: Two versions of the HTK system were used in the experiments on TREC-6 SDR test data. ... This system gave a word error rate of 24.1% on the TREC-6 test data and ran in about 50 times real time.

1 Mar 2000: On the 1998 test the system produced an average word error rate of 16.1% running in 9.5xRT. ... In tests the faster cache led to 6-8% faster decoder operation on a Xeon than on the equivalent Pentium II.

5 Oct 2000: An important issue in MMIE training is the ability to generalise to unseen test data. ... Tests were performed on both the eval97sub and the 1998 evaluation set (eval98).

1 Mar 2000: BG corr. %M corr. %Correct. %Loss. train/test. 66.26. 97.04. 97.54. 0.03. ... The test set is split into shows available both in training and test and test only.

12 Oct 2000: While MMIE is very effective at reducing training set error a key issue is generalisation to test data. ... It was found that two iterations of MMIE re-estimation gave the best test-set performance [19].

1 Mar 2000: F5. 2.7%. 2.3%. FX. 7.8%. 6.3%. Table 2: Proportion of test data of different audio type. ... The OOV rate on both these test sets using the filtered transcriptions was about 0.5%.

1 Mar 2000: For development test purposes, data broadcast in July 1996 from six shows was used. ... Since two thirds of the broadcast news training and test data is from male speakers there is a significant gender bias which isn't present in the WSJ models.

30 Mar 2000: Development. Evaluation. Name of Query Set. CU60. TREC-7 Test. Number of Queries. ... TREC-6 TEST DATA. Corr. Sub. Del. Ins. Err. NIST/IBM Baseline. 59.1.

2 Mar 2000: The experiments reported here use the TREC-7 SDR test data. For this evaluation, the audio documents came from American broadcast radio and TV news programs which had been manually divided ... The five systems described in section 4.3 were evaluated on

5 Oct 2000: All experiments were performed on the test sets used in the September 1998 and March 2000 Hub5 evaluations (eval98 and eval00, respectively).. ... It can be seen that for the eval00 test set the unmapped posteriors perform much better than for the eval98

10 May 2000: The TREC-8 SDR test collection[1] usedfor the experiments consists of 21,754 spoken documentsand 49 written queries. ... expansion. Thisparallel collection was larger and thought to be more re-liable than the automatically transcribed test collection.

30 Mar 2000: This contains the TREC-8 SDR test collection period (Feb 1st to June 30th 1998). ... For development, the 100 hours of TREC-7 SDR test data was used.

2 Nov 2000: The full HUB-4 1997 evaluation test set was used to assess language model performance. ... Treating this as the baseline result to improve upon, statistical significance was tested by using the Wilcoxon signed-rank test at a 5% significance level..

8 Mar 2000: For speech transcription problems with widely varying speaker, channel and background conditions, such as broadcast news transcription, it is beneficial to include unsupervised test-data adaptation to individual speakers and data ... The inter-node costs

2 Mar 2000: During system development the BNeval97 test set was used. 4.1 Increased Acoustic Training Data. ... All these sources excluded data from the designated test epochs. This corpus was denoted LMtrain98.

27 Sep 2000: VTLN warping in training and test both for GI and GDmodels for full and reduced bandwidth coding. ... 8. ACKNOWLEDGEMENTS. We would like to thank BBN for providing the MiniTrain trainingand MTtest test set definitions.

17 Oct 2000: standard clean modelset and then to perform further adaptation for the actual test data [7]. ... For the evaluation test H3-P0 data one microphone was selectedby NIST for each speaker.

19 Apr 2000: The windows were compared to a library of windows stored from previous shows from the broadcaster (the broadcast history) [In theory all the data in the test collection could be used ... The amount of commercial rejection with the VOA data is low, due

5 Oct 2000: 4 IMPROVING MMIE GENERALISATION. An important issue in MMIE training is the ability to generalise tounseen test data. ... The results of these experiments on both theeval97sub and eval98 test sets are shown in Table 4.

8 Mar 2000: In tests the fastercache led to 6-8% faster decoder operation on a Xeon than onthe equivalent Pentium II. ... Further tests under different operating systems indicated that. compiler efficiency was also an important issue.

8 Mar 2000: Two versions of the HTK system were used in the experimentson TREC-6 SDR test data. ... This system gave a word error rate of 24.1% on the TREC-6 test data and ran in about 50 times real time.

8 Mar 2000: approach, and two sets of HMMs were adapted.The test data for F2 was also automatically labelled as narrow. ... 1995 Hub-4 “Dry-Run” Broadcast Materials Bench-mark Tests. Proc. DARPA Speech Recognition Workshop, Har-riman, New York.

8 Mar 2000: Figure 3: Network prediction on training and test data. Actual output[.

5 Oct 2000: While MMIE is very effective at reducing training set error a keyissue is generalisation to test data. ... On other test setsimprovements greater than 1% absolute have also been obtained and.

8 Mar 2000: 9 26 142. Table 1: Training and test material available in broadcast news (min-utes) for music (M) background speaker (BGS), other background(BGO), music and speech (MS), narrow-band (T) ... The test set is split into shows available bothin training and

8 Mar 2000: news development test data andjust 15.8% on the 1997 evaluation test set. ... The OOV rate on both these test sets using the filteredtranscriptions was about 0.5%.

6 Nov 2000: ing in general) is the generalisation performance i.e. the dif-ference between training set and test set accuracy. ... Baseline Models and Hub5 Training/Test DataThree different training sets and three different test sets.

31 Jul 2000: nemea. Training data Average log2-prob (eciency). ICSI-test TIMIT-test. TIMIT 0.76! 0.60. ... derivation of the training transcrip-tions, and the hand-labelling of the test set.

31 Jul 2000: Pronunciation modeling by sharing Gaussians 141. TABLE I. WER degradation with speaking style on theMULTI-REG test set. ... Steps (6), (3), (4) and (5) arethen carried out in that order to estimate and test a matched pronunciation model.

10 Apr 2000: spontaneous non-native). Table 1: Broadcast news focus conditions. For development test purposes, data broadcast in July 1996from six shows was used. ... Overall 19.0 21.0. Table 9: % Word error rates on the Marketplace dev-test show forautomatic and

2 Nov 2000: 4 ExperimentsThe full HUB-4 1997 evaluation test set was used to as-sess language model performance.1 Although the trainingtext described in section 3 was chosen to try and matchthe ... Treating this asthe baseline result to improve upon, statistical

8 Mar 2000: During system development theBNeval97 test set was used. 4.1. Increased Acoustic Training Data. ... All these sources ex-cluded data from the designated test epochs. This corpus wasdenoted LMtrain98.

5 Oct 2000: It canbe seen that for the eval00 test set the unmapped posteriors performmuch better than for the eval98 set. ... Table 4: System Combination Results. The improvement over the single best system achieved byROVER are rather disappointing especially on

8 Mar 2000: Development work on the TREC-6 test corpus was done using twoHTK based systems. ... But more importantly, the test data is too small for reliable andinformative inference.

10 Apr 2000: Preprocess & apply SPI to the test collection to giveIt2. Preprocess & apply SPI to parallel collection to giveIp3. ... This contains the TREC-8 SDR test collectionperiod (Feb 1st to June 30th 1998).

25 Oct 2000: The test sequence(seefigure 8 ) consistsof a hollow unit cubewith textured imageson the back threefaces.This configurationwaschosenbecauseit hasa largehollow volumethathasto becarvedaway, and the exterior boundarygives no information aboutthe

10 Apr 2000: 1. INTRODUCTION. For speech transcription problems with widely varyingspeaker, channel and background conditions, such asbroadcast news transcription, it is beneficial to include un-supervised test-data adaptation to individual speakers anddata

10 Apr 2000: The output of the first pass is used alongwith a top-down covariance-based segment clusteringalgorithm [10] to group segments within each show toperform unsupervised test-set adaptation using maxi-mum ... 5. RESULTS. The five systems described in section4

6 Sep 2000: Stage2 For eachGaussianÒ we test a number(20 in this case)of randomlychosenGaussians,asin Stage1. ... Firstly, it is importantnot to test the pair. Ò<0 } any more than necessary, as calculationof overlap is time-consuming.Therefore,two optimisationsare

10 Apr 2000: The amount of commercialrejection with the VOA data is low, due mainly to the absence of any VOA broadcast history frombefore the test data. ... test collection.11 A further small increasein AveP can be obtained by increasing K.

10 Apr 2000: 5. 5.2. 5.4. 5.6. 5.8. 6. Tes. t. Time. Figure 3: Network prediction on training and test data.Actual output[ o o ] and estimated output [—]. 0 20

25 Oct 2000: 5.4 Results. The first experiment,to test the algorithm, usestwo imagesof a calibrationgrid.

10 Apr 2000: Each test segmentis allocated to the “closest” reference segment according to the distance metric being used. ... 1i.e. the speaker of the test segment must also be the speaker for at least one reference segment.

12 May 2000: Figure 11: Simulated ultrasound sweeps. (a) to (h) are the sweep patterns used to test thevolume measurement accuracy on simulated objects. ... 2% was selected as a test point since it is far enough from the resolution limit (0.7%) not to beoverly

26 Apr 2000: distance. This makes the complete test for distance initialisation as follows:. • If the voxel is not initialised, store the new distance and angle. •

10 Apr 2000: Reducing Word Error Rates of Found Speech -XPERT Tool for Transcription Analysis. S.E. JohnsonCUED/F-INFENG/TR 330. July 1998. Speech, Vision and Robotics GroupCambridge University Engineering Department. Trumpington StreetCambridge CB2 1PZ.