Nixon, J. S. & Tomaschek, F. .
Quantitative Linguistics, University of Tübingen
Within about a year from birth, infants develop from discriminating the sounds of all the world's languages to having a system more honed to their native language. It has been proposed that speech cues are learned through exposure to words and their semantic contrasts. Others argue that statistical acoustic clustering mechanisms are at work. However, lexical knowledge is initially limited and computational models suggest that statistical clustering may not be sufficient to account for speech sound acquisition. The present study neither assumes access to semantic contrasts, nor that phonetic signals are clustered on the basis of their acoustic similarity. Rather, acoustic cues are learned from error driven learning, resulting in both enhancement and decreases in phonetic sensitivity. We trained a Naive Discriminative Learner network with speech from the Karl-Eberhards-Corpus to discriminate 10 ms acoustic sequences as outcomes on the basis of surrounding acoustic input cues. Analyses indicate that acoustic outcomes form clusters, concomitant with increased sensitivity to phonetic changes on continua between contrasts. These results suggest that error-driven learning of the speech stream leads to the observed increases and decreases in sensitivity observed in infants over the first year of life before a robust vocabulary is established.