Academia.eduAcademia.edu

Figure 6 – uploaded by Dijana Petrovska-delacrétaz

See full PDFdownloadDownload figure
FIG. 6. Results for global and segmental GMM systems. Performances are reported for matched train/test conditions (SN) and mismatched train/test conditions (DT). Data from ELISA-1 (subset of NIST1998); training conditions 2s (2 min), and 3 s for test segment duration.

Figure 6 Results for global and segmental GMM systems. Performances are reported for matched train/test conditions (SN) and mismatched train/test conditions (DT). Data from ELISA-1 (subset of NIST1998); training conditions 2s (2 min), and 3 s for test segment duration.

Related Figures (5)

FIG. 1. Global and segmental speaker verification systems. Current text-independent speaker verification systems are usually based on modeling globally the probability density function of the speaker feature vectors. In such systems, denoted here as global systems, the temporal information is not taken into account and all classes are represented using a unique speaker model. As noted in the Introduction, another possibility consists of dividing the speech signal into distinct categories (also called classes or segments) and of performing the speaker modeling independently for each class. Such systems are denoted here as segmental systems. In such cases, the speaker verification task is divided into two parts: speech segmentation followed by speaker modeling for each of the classes. The general flowcharts of global and segmental speaker verification systems are shown in Fig. 1.
FIG. 1. Global and segmental speaker verification systems. Current text-independent speaker verification systems are usually based on modeling globally the probability density function of the speaker feature vectors. In such systems, denoted here as global systems, the temporal information is not taken into account and all classes are represented using a unique speaker model. As noted in the Introduction, another possibility consists of dividing the speech signal into distinct categories (also called classes or segments) and of performing the speaker modeling independently for each class. Such systems are denoted here as segmental systems. In such cases, the speaker verification task is divided into two parts: speech segmentation followed by speaker modeling for each of the classes. The general flowcharts of global and segmental speaker verification systems are shown in Fig. 1.
FIG. 2. DET curves for global GMM and global MLP systems, showing the performances for matched train/test conditions (SN) and mismatched train/test conditions (DT). Data from NIST 1998, training conditions 2F (2 min or more), and 30 s for test segment duration.
FIG. 2. DET curves for global GMM and global MLP systems, showing the performances for matched train/test conditions (SN) and mismatched train/test conditions (DT). Data from NIST 1998, training conditions 2F (2 min or more), and 30 s for test segment duration.
FIG. 3. DET curves for segmental MLP systems, showing the performances of five of eight classes, for matched train/test conditions (SN). Data from NIST 1998, training conditions 2F (2 min or more), and 30 s for test segment duration.
FIG. 3. DET curves for segmental MLP systems, showing the performances of five of eight classes, for matched train/test conditions (SN). Data from NIST 1998, training conditions 2F (2 min or more), and 30 s for test segment duration.
FIG. 4. DET curves for global and segmental MLP systems. MLPGlobC55 stands for the global system with 11 input frames and MLPSegC22RLin indicates the segmental system with linear score recombination using five input frames. Performances are reported for matched train/test conditions (SN) and mismatched train/test conditions (DT). Data from NIST1998, training conditions 2F (2 min or more), and 30 s for test segment duration.
FIG. 4. DET curves for global and segmental MLP systems. MLPGlobC55 stands for the global system with 11 input frames and MLPSegC22RLin indicates the segmental system with linear score recombination using five input frames. Performances are reported for matched train/test conditions (SN) and mismatched train/test conditions (DT). Data from NIST1998, training conditions 2F (2 min or more), and 30 s for test segment duration.
FIG. 5. DET curves for global GMM (as baseline comparison), with global and segmental MLP systems. MLPGlobC55 stands for the global system with 11 input frames and MLPSegC22RLin indicates the segmental systems (using five input frames for the MLP), with linear score recombination. Performances are reported for matched train/test conditions (SN) and mismatched train/test conditions (DT). Data from NIST1998, training conditions 2F (2 min or more), and 30 s for test segment duration.
FIG. 5. DET curves for global GMM (as baseline comparison), with global and segmental MLP systems. MLPGlobC55 stands for the global system with 11 input frames and MLPSegC22RLin indicates the segmental systems (using five input frames for the MLP), with linear score recombination. Performances are reported for matched train/test conditions (SN) and mismatched train/test conditions (DT). Data from NIST1998, training conditions 2F (2 min or more), and 30 s for test segment duration.
Related topics:
Mechanical EngineeringComputer ScienceAutomatic Speaker VerificationSpeech ProcessingSpeaker VerificationSpeech SegmentationMultilayer PerceptronElectrical and Electronic Engineering
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved