"... In PAGE 2: ... The several collections of Switchboard style corpora, each of which included hundreds of speakers and thousands of conversations, have been extensively used in the detection tasks of the NIST Speaker Recognition Evaluations. Table1 lists the corpora ... In PAGE 7: ...) Type of Transmission Training Sides Test Sides Landline 257 580 Cellular 178 361 Cordless 176 219 Other/unknown 5 16 Table 9: Phone transmission types of the training and test conversation sides for the core test condition included in the NIST 2004 evaluation data. Type of Handset Training Sides Test Sides Speakerphone 37 67 Headset 107 116 Ear-bud 42 63 Regular (hand-held) 452 914 Other/unknown 5 16 Table1 0: Phone handset types of the training and test conversation sides for the core test condition included in the NIST 2004 evaluation data. The extended data tests of previous evaluations provided (errorful) word transcripts of the speech data generated by an ASR system.... ..."

"... In PAGE 3: ...and Table3 presents their corresponding EERs. The figure and table demonstrate that the EER of the MAP-GMM is 12.... In PAGE 3: ...2. Prosody state bi-gram speaker models Figure 5 and Table3 present the performances of the prosody state bi-gram model obtained using three-segment-long super- vectors and the Good-Turing smoothing method. (In a preliminary experiment, the three-segment-long super-vector is better than the one-segment super-vector).... In PAGE 3: ... Several combinations of systems were tested. Figure 5 and Table3 present the results. The figure and table demonstrate that the EERs of the MAP-GMM and MAP-GMM+Tnorm were improved from 12.... ..."

"... In PAGE 3: ... After initialization, the parameters of the GMMs rep- resenting the claimant and the impostor speakers are up- dated with the discriminative training procedure described in Section 2. Table1 shows the costs for false alarms and false rejections, as well as the prior probabilities for claimant and impostor that we used in training (these val- ues were specified in the 1997 NIST evaluation). To pro- vide a balanced training sample, we used a 2:5 ratio of data for claimants vs.... ..."

"... In PAGE 1: ... 2. Evaluation Participants As shown in Table1 , the participants over the past six years have been from 12 countries on five continents, making these truly worldwide evaluations. In some cases two or more participants have worked in cooperation while submitting individual results for separate systems.... ..."

"... In PAGE 1: ... 2. Evaluation Participants As shown in Table1 , the participants over the past six years have been from 12 countries on five continents, making these truly worldwide evaluations. In some cases two or more participants have worked in cooperation while submitting individual results for separate systems.... ..."

"... In PAGE 2: ... Experiment results Several speaker recognition experiments were performed to evaluate the various feature extraction methods. For speaker recognition experiments, the results for three meth- ods are reported in Table1 . The best recognition rate of AFCM WT obtained was 95%.... ..."

"... In PAGE 4: ...2 Table 2: Performance of discriminative trainingprocedure on 1997 NIST Evaluation (male, 10 second test, 1 session training) for one speaker on training and cross validation sets. Table3 shows results for the 1-session (2 minutes of training from one phone call), 10-second male test in the 1997 NIST Speaker Recognition evaluation where the same telephone was used in by the claimant in training and testing. Table 4 shows results for the same training case, but with tests from a different telephone than used in training.... ..."

"... In PAGE 7: ...) Type of Transmission Training Sides Test Sides Landline 257 580 Cellular 178 361 Cordless 176 219 Other/unknown 5 16 Table 9: Phone transmission types of the training and test conversation sides for the core test condition included in the NIST 2004 evaluation data. Type of Handset Training Sides Test Sides Speakerphone 37 67 Headset 107 116 Ear-bud 42 63 Regular (hand-held) 452 914 Other/unknown 5 16 Table1 0: Phone handset types of the training and test conversation sides for the core test condition included in the NIST 2004 evaluation data. The extended data tests of previous evaluations provided (errorful) word transcripts of the speech data generated by an ASR system.... ..."

"... In PAGE 3: ...1. Speech Material and Parameter Setup For the experiments, we used a subset of the NIST 1999 speaker recognition evaluation corpus [21] (see Table2 ). We selected to use the data from the male speakers only.... ..."

"... In PAGE 2: ...1. Automatic prosody state labeling and bi-gram speaker models Clustering the extracted prosodic super-vectors of all registered speakers enabled eight- and three-codeword VQs (see Table1 ) for voiced and unvoiced segments, respectively, to be learned and used to model the prosodic characteristics of all speakers. In particular, the trained VQs were used to label automatically the input sequences of an input utterance into sequences of prosodic states.... ..."