Compared to low-variability training, high-variability training leads to better learning outcomes and supports generalization of learning. However, it is unclear whether the learning advantage is driven by multiple talkers or by enhanced acoustic variability across target sounds. The current study addressed this issue in nonnative production learning. Spanish speakers were trained to produce the French /e/--/?/ vowel contrast. The stimuli were recorded by five native French talkers for the multiple-talker (MT) group or by one talker for the single-talker (ST) group, but acoustic dispersion of the vowels and context were matched between the two groups. Both training paradigms improved production accuracy, with slightly greater improvement in the ST group. However, only MT training enhanced the compactness of vowel categories and generalized to the production of sounds elicited by an unfamiliar speaker. This suggests that talker variability supports the establishment of abstract phonemic categories in production.