This study investigated repetition priming in the recognition of famous voices, recording reaction times (RTs) and event-related brain potentials (ERPs). In Experiment 1, a facilitation was found in RTs to famous but not to unfamiliar voices when these had been primed by a different voice sample of the same speaker earlier in the experiment. However, ERPs to both famous and unfamiliar voices showed repetition priming in terms of an increased P2 component, which is thought to be generated in the auditory cortex. When the likelihood of conscious retrieval of primes was reduced in Experiment 2, facilitatory priming in RTs was again observed for famous voices, but inhibitory priming was now observed for unfamiliar voices. This is consistent with predictions of a bias model of priming. Moreover, substantial priming was observed even when voice primes were backward speech samples, which were recognised at chance levels. The results suggests that (a) voice priming is mediated to a large extent by frequency characteristics of a particular voice, rather than by articulatory and other 'sequential' features that are eliminated in backward speech; (b) priming affects the processing of voices in auditory cortical areas within 200 ms after voice onset; and (c) explicit recognition of a voice in the priming phase is not a necessary condition for priming to occur.