|Addresses: www mail|
According to the ecological and emergent approach to language learning, infants are able to learn word-like patterns without preprogrammed linguistic knowledge such as phonemes. It is believed that the first words are learned by using relatively simple pattern matching techniques. Instead phonemes emerge as the vocabulary grows and statistical models are needed to handle the increasing complexity.
This work shows how pattern matching techniques can be used to create an initial set of words through the natural interaction between an infant and its caregiver. It also shows how a statistical "phoneme" model can emerge from this initial set of words in an unsupervised way. The learning techniques are implemented and demonstrated on a humanoid robot.