A more diagnostic approach to compiling a test database has been proposed by [Peckham & Thomas (1990)]. Here, an Analysis of Variance method (ANOVA) is used in order to measure the influence of specific parameters of the input signal to the recognition score.
A number of parameters that can influence the recognition score are defined; these are typically speech parameters such as speaking rate , fundamental frequency , language parameters such as perplexity , sentence length, and quality parameters such as signal-to-noise ratiosignal-to-noise ratio . For each of the parameters whose influence is to be measured, a number of ``levels'' is defined, e.g. ``high'' and ``low'' (perplexity ) or ``0-10'', ``10-25'' and ``25-50'' (dB signal-to-noise ratio ). The database then should cover all cells of the design with a sufficient number of observations.
Because the number of parameters M readily increases the total number of cell s, M is practically limited. Some of the parameters may be uncontrolled (speaking rate ), others may be controlled (signal-to-noise ratio ). For uncontrolled parameters it can turn out that not all cells are filled with enough observations; in that case one may have to reduce the number of parameters in the analysis.
When all of the parameters in the database have been categorised into groups, the database is called calibrated.