Determining Learner Pronunciation Models with the K-nn Algorithm

Most phonetic learner corpora are dictionary-based and transcribe a target / pronunciation model for learners [1], which is usually an American pronunciation model, because the electronic resources for British English are scarce. This paper proposes to investigate which reference pronunciation model learners aim to emulate, by comparing their realisations with two reference formant datasets, which represent pronunciation models for British and American speakers. Reference varieties (and their putative rejection or adoption in the name of ELF) have been the topic of a heated debate, sometimes laden with political and economical considerations ([9], [12]). The whole debate ([5], [7], [10], [12]) and the economic issues surrounding prestige [9], as well as ideological and political implications have been commented upon, as well as the potential need for a “Lingua Franca Core” [12]. Various methods have been used to determine a variety of English on segmental criteria. Wieling et al. [13] have adopted another methodology to “evaluate the suitability of a computational pronunciation comparison method” and have used ready-made transcriptions from the ACCENT project archives and have shown the Levenshtein distance to be a good metrics for the measure of pronunciation distance, congruent with on-line native judgements.