Co-evolution of language acquisition and infant-directed speech Bart de Boer Vrije Universiteit Brussel bartb@arti.vub.ac.be This abstract presents an investigation of the role of infant-directed speech in the evolution and acquisition of complex language. It is proposed that infant-directed speech (the adaptation of a caretaker's utterances to the language skills of an infant) helps to stabilize more complex sound systems in a population. This enables more complex language to persist in a population and thus creates evolutionary pressure for better language acquisition skills. These ideas are investigated with a computer model and observations of real infant-directed speech. The most frequent speech used by adults is rapid, casual speech. The utterances, and especially the vowels in this speech register tend to be strongly reduced. If infants would therefore be purely statistical learners, they would tend to learn reduced versions of the sound system of a language and especially of its vowel system. As this process is repeated over the generations, sound systems would rapidly collapse. This is not observed in practice, however. Sound systems, although changing over time, do not collapse. It must therefore be concluded that infants are not purely statistical learners. Learning systems can deviate from purely statistical learning in two ways. Either the learned categories are manipulated to compensate for reduction, or learners make a selection of the input data and base their learning on the selected examples. The first hypothesis (the compensation case) can be defended by the fact that children probably have to perform a number of compensations when learning speech anyway. The adult vocal tract is different from the infant vocal tract and in order to learn the sounds of a language, a child has to compensate for this difference. An added compensation for reduced speech is then not unlikely. The second hypothesis (the infant-directed speech case) can be defended by the existence of infant-directed speech. Such speech appears cross-culturally and has properties that make it easier to learn (slower tempo, more exaggerated intonation and articulation, face-to-face interaction etc.) If infants would base their learning on infant-directed speech preferably, reduction would not be expected. Transfer of vowel systems was implemented in an agent-based computer simulation. A population of agents consists of adults and infants. Adults have a fixed vowel system, and produce noisy, reduced utterances. Infants learn a vowel system from adult utterances. After a number of interactions, adult agents are removed from the population, infants fix their vowel systems and are converted into adults, and new, empty infant agents are inserted in the population. In the compensation case infants listen to reduced adult utterances, and compensate for reduction by expanding the vowel systems they learn. In the infant-directed speech case, adult agents produce less reduced utterances. It is found that for five-vowel systems, both mechanisms transfer vowel systems perfectly. However, for larger (seven) vowel systems neither of them works well. In combination, however, they do preserve vowel systems in the population. As compensation can be argued to exist on independent grounds, the conclusion is that infant-directed speech is needed as an extra mechanism to preserve complex vowel systems. This conclusion is supported by observations of real infant-directed speech. It is found that in languages with more vowels, infant-directed speech is more pronounced than in languages with fewer vowels. The implication for language evolution is that relatively simple behavior by adults can facilitate the task of learning language by infants. This makes it possible for more complex language to persist in a population. If more complex language is stable in cultural transfer, this makes it advantageous for agents to evolve adaptations for such complexity. Thus adult adaptation to infant learning can bootstrap the increasing complexity of language.