Performance but Also on the Ability of the System to Reject an Unknown User. Evaluation of a Vector Quan- Tization Talker Recognition System in Text Independent And

7 the so called Min/Max ratio (see 15], 16]) which is deened as the minimum distance on a wrong correspondence over the maximum distance from the correct correspondence. To estimate the optimal weighting factor the interval 0; 1], representing the weight of the FRS, was evenly sampled and the value maximizing the performance was chosen (ties were resolved by maximizing the Min/Max ratio). The performance of the SRS, FRS and integrated systems are quantiied by: 1. the average performance (Figure 6); 2. the Min/Max ratio (Figure 7). The analysis of this integrated performance (see Figure 6) shows that, even with the simple integration scheme, nearly perfect recognition is achieved. The performance of the FRS is 92% while that of the SRS, using diierent speaker models, varies from 51% to 100%: the performance of the integrated system, using the less complex speaker model is already up to 95% and achieves 100% with a speaker model of low complexity. Further beneets of integration are evident in Figure 7, where the Min/Max ratio, which represents an estimate of the average separation among classes, is plotted. The increased class separation of the integrated system, as measured by the Min/Max ratio, suggests that rejection can be introduced with a more limited impact on performance if compared to the two independent systems. The curves in Figure 8 represent the performance of the integrated system as a function of the weight used to merge the scores; as can be seen the performance is a smooth function of the weight: this means that the the system is not very sensitive to the weighting factor. As it could be easily anticipated from the performance of the two diierent systems, the optimal combining weight shifts from a vision dominance at a low number of digits toward a voice dominance at an high number of digits. The same result would be reached passing from lower to higher spectral resolutions of the SRS.