Support vector machines versus fast scoring in the low-dimensional total variability space for speaker verification

摘要: This paper presents a new speaker verification system architecture based on Joint Factor Analysis (JFA) as feature extractor. In this modeling, the JFA is used to define low-dimensional space named total variability factor space, instead of both channel and spaces for classical JFA. The main contribution in approach, use cosine kernel design two different systems: first Support Vector Machines based, second one uses directly decision score. last scoring method makes process faster less computation complex compared others methods. We tested several intersession compensation methods factors, we found that combination Linear Discriminate Within Class Covariance Normalization achieved best performance. remarkable results using fast only especially male trials, yield an EER 1.12% MinDCF 0.0094 English trials NIST 2008 SRE dataset. Index Terms: Total kernel, scoring, support vector machines.