Talkers vary in how they speak, resulting in acoustic variability in segments and prosody. Previous studies showed that listeners deal with segmental variability through perceptual learning and that these learning effects are stable over time. The present study examined whether this is also true for lexical stress variability. Listeners heard Dutch minimal pairs (e.g., VOORnaam vs. voorNAAM, ‘first name’ vs. ‘respectable’) spoken by two talkers. Half of the participants heard Talker 1 using only F0 to signal lexical stress and Talker 2 using only intensity. The other half heard the reverse. After a learning phase, participants were tested on words spoken by these talkers with conflicting stress cues (‘mixed items’; e.g., Talker 1 saying voornaam with F0 signaling initial stress and intensity signaling final stress). We found that, despite the conflicting cues, listeners perceived these items following what they had learned. For example, participants hearing the example mixed item described above who had learned that Talker 1 used F0 perceived initial stress (VOORnaam) but those who had learned that Talker 1 used intensity perceived final stress (voorNAAM). Crucially, this result was still present in a delayed test phase, showing that talker-specific learning about lexical stress is stable over time.