Perceptual learning about voices is known to facilitate speech perception, but it is unclear exactly which phonetic representations are altered to cause this facilitation. This study examines perceptual learning for a non-segmental phonetic property, talker-specific cues to word boundaries. An experiment tested intelligibility in noise of sentences that contained hard-to-segment sequences (e.g. /patsɔːd/, which can correspond to Pat sawed or Pat’s awed). Testing occurred before and after training with a voice; improvement in performance after training was measured. Subjects who heard the same voice during training as during testing showed more improvement than those who heard a different voice. Implications for exemplar theories of speech perception are discussed.