Modeling L1/L2 interactions in the perception and production of English vowels by Mandarin L1 speakers: A training study

This dissertation proposes a new statistical pattern recognition model for measuring crosslinguistic vowel similarity and applies it to measuring differences between English and Mandarin vowels. The model incorporates multidimensional acoustic information previously found to be important for vowel identification. Previous research has trained similar statistical models on vowel production values from a single language and used these trained models to determine how new production tokens from a second language might be classified in terms of the first language. The current statistical model was trained on production values from both languages being compared, Mandarin and English. New production tokens from both languages were then tested against this statistical model to determine the extent to which production tokens in one language were misidentified as members of opposing language categories. The degree of misidentification across languages provides a metric for determining how similar Mandarin and English vowels are to each other. From this, explicit predictions were made concerning how Mandarin speakers would identify and produce L2 English vowels. In a training study, twenty-six L2 learners were trained under three conditions to identify ten English vowels. Changes in the learners' ability to identify and produce these English vowels were used to measure the effect of instruction as well as predictions of the statistical model of Mandarin-English vowel similarity. Results indicate that this statistical approach to measuring crosslinguistic similarity can be used to quite accurately predict the behaviour of adult Mandarin learners of L2 English. Furthermore, the results demonstrate that under certain training conditions, Mandarin learners of English are able to improve in their ability to identify and produce English vowels.