In Automatic Speech Recognition it is rather straightforward that large and diverse spoken language databases are required to train and to test speech input systems. For unlimited-text-to-speech rule-synthesis systems this is less apparent. These systems speak with one or few voices only, cannot just imitate natural speech, apply rules that are a mixture of database 'facts', designer's intuitions, and compensations for system insufficiencies, whereas subjective evaluation requires different texts all the time. Still, carefully designed, relatively small, annotated, spoken language databases can be an indispensable source of information and can be of some value as a standard for comparison. Such databases should be appropriately 'documented' in terms of orthographic input, grammatical categories, prosodic labeling, detailed acoustic-phonetic labeling, and voice characteristics.