ABSTRACT
Peoples' voices naturally appear to be a central component of their identities but under experimental conditions speaker recognition performance by humans is variable and often extremely poor. Reported accuracy rates have varied from 96% correct (McGehee, 1937) to chance (e.g., Saslove & Yarmey, 1980). Near chance performance has been reported even when the target speakers were familiar to the listener, if the listener only heard one word, although performance improved to 90% correct after 33 words (Goldstein & Chance, 1985). In reviews, Bull and Clifford (1984; Clifford, 1983) concluded that evidence to that date suggested that human recognition of speakers is too unreliable to be forensically useful for person identification. Since then, researchers have tried harder to simulate natural conversational conditions but some unsolved problems remain. In particular, research has failed to provide a coherent account of why experimental speaker recognition performance varies so much. This chapter provides a tentative answer to that question. It suggests that it is unlikely that people can reliably identify by voice strangers whom they have heard once briefly. For, as is described, people's voices vary too much from occasion to occasion and a single brief exposure may fail to provide sufficient voice information for adequate transfer to another occasion. Almost all studies of speaker identification by humans have involved the recognition of strangers heard briefly once, so it is not surprising that only a few have found significant levels of recognition,
