What is the ideal test population like? It should be representative (cf. Chapters 4 and 9).
There are two ways of collecting representative test populations:
The advantage of the latter strategy is that the number of subjects to be tested can be reduced drastically in comparison with the purely statistical approach. Furthermore, it is always possible to select a specific subset from the whole population in order to focus the investigations on specific problems. This has to be traded off by some extra investment in the clarification of the set of meaningful descriptors as well as by extra effort to be spent on the classification of each subject according to these descriptors.
For the statistical approach, however, there is no need for anything like an explicit classification; this may be crucial for on-site clandestine speech recordings.
Independently of the purpose for which a population is selected, listeners or talkers, its descriptors may be divided into two types: general descriptors and task specific descriptors.
For the reasoning concerning which descriptor to select and why, consult Section 3.5. Here we focus on how to extract the relevant physical characteristics.
It is important to note that most of the physical descriptors, e.g. voice quality or precision of articulation, are long-term functions of time. Consequently, test populations are, generally speaking, subject to aging.