Corpus-based speech synthesis needs representative corpora of human speech if it is to meet the needs of everyday spoken interaction. This paper describes methods for recording such corpora, and details some difficulties (with their solutions) found in the use of spontaneous speech data for synthesis.