This paper describes an investigation of two computer models of how vowel systems can be transferred from one generation to the next. Humans tend to reduce the articulation of the vowels (and other speech sounds) they produce. If infants would learn on the basis of these reduced signals, vowel systems would collapse rapidly over time. As this is not observed in practice, some mechanism must be present to counter it. Two candidate mechanisms are investigated in this paper: compensatory expansion of articulations learned on the basis of reduced speech sounds and learning on the basis of more carefully articulated speech. It turns out that larger vowel systems with central vowels can only remain stable when learning is based on carefully articulated speech.