"... In PAGE 4: ...3. Baseline system comparison and combination The top part of Table5 gives complete results for our two cep- stral baseline systems, as well as the MLLR systems using 2+2 or 8+8 transforms. We observe that the results across all data sets are quite consistent, and, in particular, SRE-05 results are very similar to those on SRE-04.... In PAGE 4: ... Interestingly, the 2+2-transform MLLR system is competitive with the MFCC GMM system, and beats it in the 8-side condition. The middle and bottom parts of Table5 shows results with combinations of the two MFCC baseline systems with the MLLR systems, using a neural network for combining the sys- tem output scores. The combiner is trained to minimize DCF on the SRE-04 data sets.... In PAGE 4: ... A three-way com- bination does, however, improve over the best two-way system, yielding 22% (for 1-side training) and 10% (for 8-side training) relative EER reduction over the MLLR system by itself. The middle row in Table5 shows that even the 2+2- transform MLLR systems can boost the accuracy of a GMM baseline system signficantly when combine with the latter. This might be of interest if full word recognition is not an option, as... ..."

"... In PAGE 4: ...The gender-dependent performance results, for the two systems shown in the Table1 , were computed at the EER decision point. The expected superiority of the SpkVDep performance, in comparison with the SpkVUni one, was not observed.... In PAGE 4: ... On the other hand, the idea of using speaker-dependent background codebooks is useful in closed-set scenario, when only a limited number of enrolled speakers (and/or a small number of non-users) are available for constructing the background models. The gap between the performance of the male and fe- male speakers (Figure 2 and Table1 ) could be explained by the traditional difficulties with the female speech the speaker verification systems have, in opposition to the human listeners which have difficulties with the male voices [13]. Individual speech habits are essential cues for speaker verification performed by humans, and there- fore, factors like: pronunciation, prosodic style and word choice, should be taken into account when the speaker models for ASV are build.... ..."

"... In PAGE 2: ... Results are given in Tables 3 and 4. In Table3 , line (1) corresponds to the base- line (TIMIT EFR experiment reported from Tables 1 and 2). When extracting features from encoded parameters, we have a frame rate (imposed by the EFR coder) of 20 ms.... In PAGE 3: ... 5.2 Use of Higher Order LPC In Table3 it is observed that increasing the LPC order improves the performance, but only 10-th order LPC is available in the EFR encoded parameters. Different experiments we have carried out let us assume that higher order LPC information leaks in other encoded parameters (LTP lags and gain, and stochastic pulses and gain) and is thus available in the decoded speech, improving recognition.... In PAGE 3: ... We investigated the use of this higher order LPC informa- tion. The goal is to improve upon (14) in Table3 , the best... In PAGE 4: ... Naturally, the performance is still better when extracting features from the original speech. We have improved upon line (14) in Table3 , got close to the baseline for speaker identification and improved upon the baseline for verification. 6.... In PAGE 4: ... The performance is also improved by using LSP parameters instead of cepstral coefficients. The best result we have obtained (line 14 in Table3 ) is slightly worse than the baseline in performance (line 1 in Table 3), but computationally more efficient (amount of feature vectors is halved, and vector dimension is reduced from 16 to 11). Future work should include finding ways of improving the baseline, varying either the speaker recognition system, or the feature extraction.... In PAGE 4: ... The performance is also improved by using LSP parameters instead of cepstral coefficients. The best result we have obtained (line 14 in Table 3) is slightly worse than the baseline in performance (line 1 in Table3 ), but computationally more efficient (amount of feature vectors is halved, and vector dimension is reduced from 16 to 11). Future work should include finding ways of improving the baseline, varying either the speaker recognition system, or the feature extraction.... ..."

"... In PAGE 3: ... Mismatched channel conditions were simulated using a) speech recorded through a PSTN, b) speech recorded through a cellular-telephone (GSM) and c) speech recorded in the calling room (Room). Figure 2 and Table1 show the equal error rates (EERs) of a classical GMM-UBM based speaker verification system compared to the Bayesian network system. Table 1: EERs when the training data is speech recorded through PSTN... ..."

"... In PAGE 4: ... 5. Results Table3 lists two sets of experimental results. The two experiments are different in the way the phoneme alignments for CPM were ob- tained.... In PAGE 4: ... The two experiments are different in the way the phoneme alignments for CPM were ob- tained. One set of results is summarized under Recognized Align- ment and the other under Forced Alignment in Table3 . The table is divided into three columns: Matched , Mismatched and All .... In PAGE 4: ... This was achieved by forced aligning the phoneme sequences of all enrollment and verification utterances with the transcribed word sequences and lexicon obtained from [8]. The results of using forced phoneme alignments are summarized under the heading Forced Alignment in Table3 . The overall EER is reduced to 22.... In PAGE 4: ...educed to 22.69%. The reduction from 25.83% to 22.69% suggests that the accuracy of phoneme alignments is critical to the verifica- tion performance of the AFCPM system. The experimental results of the MFCC system and the fusion systems are also summarized in Table3 . The fusion weights of the systems were determined from a four-fold cross validation using the test data from target speakers and non-target speakers.... In PAGE 4: ... The results demonstrate the effectiveness of adjusting the contribution of individual frames according to the score confidence in each frame. From the results corresponding to matched handsets and mismatched handsets in Table3 , it is clear that fusion of the spectral-based scores and AFCPM scores plays an important role in reducing the EER under handset mismatched conditions. As the spectral features are less reliable under handset mismatched condi- tions, the speaker information from AFCPM becomes more impor- tant.... ..."

"... In PAGE 6: ....6.1. Forced Face Detection Results Table1 shows our user verification results for three systems (face ID only, speaker ID only, and our full multi-modal sys- tem) under two different impostor conditions (using only known in-set impostors vs. using only unknown out-of- set impostors).... ..."

"... In PAGE 6: ....6.1. Forced Face Detection Results Table1 shows our user verification results for three systems (face ID only, speaker ID only, and our full multi-modal sys- tem) under two different impostor conditions (using only known in-set impostors vs. using only unknown out-of- set impostors).... ..."

"... In PAGE 10: ...2 Baseline Results In this section, we present two different baseline results. First in Table1 , we show the performance of each modality alone, namely speaker verification (Voice) and face verification (Face). For both, we give results for configurations I and II, and for each configuration, FAR represents the false acceptance rate, FRR is the false rejection rate, while HTER is the half total error rate.... ..."