Diagram 2.5 establishes the distribution of errors across speakers. It indicates the quality of the test database , and shows that there may be fewer than 2 errors for 99% of the speakers (``sheep speakers'') and about 10 for some ``goat speakers''.
The two approaches described here (confusion matrix and speaker selection) may be influenced by the fact that building databases can be biased, and should be carried out carefully.
Figure 2.5: Result presentation with ``speaker selection'' procedure