This paper describes a corpus for the study of intraspeaker variability and of contrastive speaking styles. Design accounted for: different speaking styles being compared for the same linguistic content for a given speaker; the speaking styles being clearly produced and perceived to be different without direct prompting or reading; a series of specific predetermined contexts of phonological variation being elicited. In the "seven errors" task, the speaker describes objects which differ in two drawings and are places where possible phonological variation may occur. Styles include casual, clear and clear read. The signal from one speaker occupies about 60 MBytes (audio) in all, about 30 minutes of speech, representing an average of about 70 "phonological contexts" per speaker per style, covering: voicing, devoicing, schwa elimination, palatalisation, nasalisation, and geminates. Twenty one speakers of varying origins have already been recorded. The data is being labelled, using semi-automatic labelling techniques and will be perceptually verified to confirm style change. This methodology has also been used to collect a corresponding corpus of British English.

Keywords: speaking styles, variability, speaker characterisation, database