Search

20 Dec 2013: The traditional approach to assessing spoken English is to have a well-trained human assessor listen to the test - either live or recorded - and mark the performance on a standardised scale.

13 Jun 2013: As the test data was Levantine Arabic,there is little appropriate data for training language models. ... Forthis work all the clean Levantine Arabic transcriptions, exclud-ing the dev1 test data, was used.

13 Jun 2013: Part of theretransmitted data was held-out to form a test set, dev1. ... Foreach of the channels there was 2 to 2.5 hours test data, de-pending on how much of the retransmitting speech passedquality assurance tests.

13 Jun 2013: noisesource (“restaurant noise”), while the 410 test utterances weredistorted by the 6 noise sources with a uniform distribution. ... The first row of table 1 shows the perfor-mance on three test sets.

13 Jun 2013: The recognition task is a 5K-word dictationtask with 14 test sets, 330 utterances each. ... In the follow-ing, letters A, B, C, and D will be used to indicate test sets 01,02-07, 08, 09-14, respectively.

13 Jun 2013: All the experts (SVMs) of the iSVM share the same C,and the parameter C is tuned on the test set A. ... Figure 5: The performance of the iSVM on test set A with dif-ferent C.

7 Nov 2013: Tandem systems. A language independent acoustic model test on. the target language showed that retraining or adapting of the acous-. ... used for alignments. At recognition time a language specific LM,. trained on the transcriptions from the language

13 Jun 2013: This dataset wasfurther split into a training set for the CRF models, and a test setfor evaluation. ... Table 1. Results for the word and grapheme-based KWS systems ontraining and test subsets of dev-1, with no threshold applied to con-fidence scores.

13 Jun 2013: For the test sets, the330 utterance from 8 speakers in the AURORA4 set A were fil-tered by two RIRs, “office1” and “office2”, where the latter wasrecored in another office environment ... Recognition in a noisy andreverberant environment is the

27 Mar 2013: The maximum word accuracy criterion was used to train dis-criminative models on multi-style data using suboptimal alignments.Test set A was used as the validation set to stop training. ... However, in this case the SDM givesimprovement on test set B only.

27 Mar 2013: Test setA was used as the development set for tuning parameters forall systems, such as the C in (4). ... SNR Test Set A(dB) HMM M-SVM LLM -1 LLM-2 LLM-320 1.7 1.5 1.4 1.3 1.115 2.4 2.0 1.9

13 Jun 2013: 129–132. [2] “NIST spoken term dection portal,” http://www.itl.nist.gov/iad/mig/tests/std/. [3] “The 2006 spoken term detection evaluation ... mig/tests/std/2006/docs/std06-evalplan-v10.pdf.

9 Sep 2013: ese scores then form features for the discriminative classiers. is. Test data O. ... is makes it possible to test the interaction with noise compensa-tion methods.

27 Mar 2013: 4. EXPERIMENTS. The feature extraction process described in this paper was tested ina log-linear model on a small, noise-corrupted corpus: AURORA 2.This makes it possible to test the ... Test setSNR A B C(dB) HMM l l,l HMM l l,l HMM l l,l20 1.69 1.43

13 Jun 2013: It contains 200 hours oftraining audio; approximately 50% of the training data is silence.The test data is limited to only conversational data.

27 Mar 2013: is makes it possible to test the interaction with noise compensation methods.e task uses a small vocabulary and no language model, which makes experimentswithout such optimisations as pruning possible. ... One of the three test sets, test set A,was used