Bond, S.
Oxford Brookes University
The ability to extract statistical information from speech is an important part of language learning, helping to locate word boundaries and acquire governing rules, like grammar. In phonological learning some statistical cues appear more important than others in determining acquired phonotactics, indicating underlying constraints on our statistical learning mechanism. The studies presented here explore these constraints using artificial language, allowing for complete control of the statistical characteristics in presented phonetic sounds. The artificial languages developed in these experiments had a rigid hierarchy of statistical cues: co-occurring consonants gave word-level regularities, yet overarching single-feature (voicing) and multiple-feature (voicing and manner of articulation) phonological biases were also present. 144 participants heard three blocks of continuous speech in a 21minute familiarisation phase. Participants were asked to identify legal words in 24 test pairings before completing a word search task, measuring word productions based on one of the three levels of statistical cue. Data analysis focussed on whether participants showed a preference for statistical information occurring at word-level or for the single-feature or multiple-feature phonological biases. Results discussed rule generalisation and any apparent level preference, as well as differences between manners of articulation. Opportunities for future studies are raised.