EVOLUTION OF A RAPIDLY LEARNED REPRESENTATION FOR SPEECH

Newly born infants are able to finely discriminate almost all human sp eech contrasts and their phonemic category boundaries are initially id entical, even for phonemes outside their target language. A connection ist model is described, which accounts for this ability. The approach taken has been to develop a model of innately guided learning in which an artificial neural network (ANN) is stored in a ''genome'' which en codes its architecture and learning rules. The space of possible ANNs is searched with a genetic algorithm for networks that can learn to di scriminate human speech sounds. These networks perform equally well wh en they have been trained on speech spectra from any human language so far tested (English, Cantonese, Swahili, Farsi, Czech, Hindi, Hungari an, Korean, Polish, Russian, Slovak, Spanish, Ukranian, and Urdu). Tra ining the evolved networks requires exposure to just two minutes of sp eech in any of these languages. Categorisation of speech sounds based on the network representations shows the hallmarks of categorical perc eption. Phoneme confusability in the network replicates earlier studie s of phoneme confusability in adults. The network model offers an epig enetic account of the rapid emergence of speech perception skills in y oung infants whereby innately specified neural systems exploit regular ities in the speech signal to construct representations that are well- suited to the identification of speech segments. The model also sugges ts how infants' early preferential attention to speech is driven by th e rapid construction of suitable representations.