It has been claimed that corpus-based TTS is unworkable because it is not practical to include representative units to cover all or most of the combinations of segments and prosodic characteristics found in general texts, a problem characterized as Large Numbers of Rare Events (LNRE). We argue that part of this problem is in its formulation, and that a closer look, including investigations into corpus-based TTS for Danish, show that LNRE need not be a fatal problem for inventory design in corpus-based TTS.