Speech the Value of a Longitudinal Corpus

This paper describes a new corpus that has been created for speech technology research. It includes the voices of several speakers recorded over an extended period of time, and illustrates how speaking style and voice quality can change according to differences in interlocutor and utterance content. Whereas most corpora are purpose-designed and collected under controlled circumstances, often resulting in a constrained style of speaking, the ESP corpus (from the JST/CREST Expressive Speech Processing project) was collected to illustrate the wide range of speaking-styles that can occur in ordinary everyday conversational situations. It is limited in the number of speakers and contexts of the dialogues, but provides a unique sample of everyday conversation for the speech researcher.