Frequent or contextually-predictable words are often phonetically reduced, e.g. shortened or produced with articulatory undershoot. Three common explanations for this phenomenon attribute phonetic reduction, and pronunciation variation generally, to variation in (1) intelligibility, (2) speed of lexical access, and (3) probabilistic properties of whole utterances. In this talk, I discuss recent results (Gahl, Yao & Johnson, under review; Gahl, in prep.) investigating the sources of pronunciation variation. While these results are consistent with speaker-internal approaches to variation, I argue that they are best explained by moving beyond the listener vs. speaker dichotomy: Some perceptually-based effects are speaker-internal; and some production-based effects give rise to structured variation in intelligibility resulting in usable cues for recognition.