This paper proposes a paradigm where commonly made segmental pronunciation errors are modeled as pair-wise confusions between two or more phonemes in the language that is being learnt. The method uses an ensemble of support vector machine classifiers with time varying Mel frequency cepstral features to distinguish between several pairs of phonemes. These classifiers are then applied to classify the phonemes uttered by second language learners. Using this method, an assessment is made regarding the typical pronunciation problems that students learning Swedish would encounter, depending on their first language.
tmh_import_11_12_14. QC 20111220. QC 20120111