Confidence analysis for text-independent speaker identification: inspecting the effect of population size

Confidence analysis was incorporated into a text-independent speaker identification procedure to get the confidence with which one can state the target speaker's presence in or absence from the reference list. The degree of confidence was deduced from a classification of two groups of curves, namely AvD (accuracy versus distance) and PAvD (pseudo accuracy versus distance). The authors inspect the effect of population size (ranging from 5 to 19) on confidence analysis. The inspection is performed by the Fisher's discriminant on the distribution of the classification results of AvD and PAvD curves. Experiment shows that confidence analysis can improve only very little as the population size decreases. Other factors (e.g., interspeakers similarity) seem more dominant than population size.<<ETX>>