The ability to perceive sounds as words involves a transformation from detailed speech signals to invariant meanings, which are separate from information about the speaker of a particular word. The nature of this transformation is a central issue in the field of speech perception. A particular focus of ongoing debate concerns talker-specific details: are they causally relevant to lexical perception, or are they useful only for tasks like speaker recognition? One common way to investigate the impact of voice information is to examine the time-course of its effects on future perceptual events. Early research reported no consistent long-lasting effects, implying that speech representations do not contain talker-specific detail (Jackson & Morton, 1984). However, subsequent work reported long-lasting effects, leading to a focus on modelling speech representations as abstractions over detail-rich episodic memories (Goldinger, 1996). Current hybrid models (Church & Schacter, 1994; McLennan & Luce, 2005; Goldinger, 2007) incorporate abstract and detail-rich speech representations but differ in the relative importance assigned each. Two types of hybrid models are differentiated: a) models with combined representations, where abstraction occurs over detailed memories of speech episodes; versus b) models with separate representations, where different processing paths exist from the speech signal to word and speaker recognition. To investigate these models, this thesis reports multiple experiments investigating the time-course of the decay patterns of voice effects in repetition priming. Results from auditory lexical decision indicate that voice information only affects the speed of future perceptual processes within a short time window: until around three items intervene between prime and target. This finding clarifies previous results, which found no long-lasting effects, by providing an exact time-course of voice information’s impact. Nevertheless, the results reported here differ from the predictions of studies investigating recognition accuracy, where long-lasting effects are commonly found. To address these differences, additional experiments using continuous and blocked word recognition paradigms were conducted. Again, talker-specific effects only persist within the same short time window, while abstract repetition priming effects persist much longer. By de-emphasizing the contribution of voice information, these findings assert the importance of abstract linguistic representations in hybrid models with separate representations.