Rather than collecting everything we can try to collect only the ``right'' data. There are many suggestions of designing database inventory and utterances that cover the desired space. For example  use an elaborate scheme where they first model the acoustic space of a speaker, thus finding out which units are acoustically distinct, and frequent enough to deserve coverage. The second stage, more conventionally, greedily selects utterances from a database to best cover that required inventory. The result is a reasonably manageable set of utterances (perhaps 500-1000) which covers the identified acoustic/phonetic space well. Such techniques give better synthesis (per database size) than non-designed databases.