Phrase accentuation verification and phonetic variation measurement for the degree of nativeness sub-challenge

The Degree of Nativeness Sub-Challenge consists in the automatic grading of the pronunciation quality of non-native English utterances. In this paper, we investigate the phrase accentuation and the phonetic acoustic variability for the prediction of the grades. Two prediction systems have been developed: the Extended Baseline System (EBS) and the Pronunciation Feature based System (PFS). The EBS system was designed to take into account the cross-corpus specificities such as recording conditions and the sentence variability. The speech files were segmented using Automatic Speech Recognition methods (ASR). Audio features were selected on both the training and development sets using the Regressional ReliefF method. New audio features were developed for the PFS system to take into account the mispronunciations: unusual prosody and/or phonetic variation. These systems have been assessed using the Spearman’s correlation coefficient with expert annotations. The PFS system has significantly improved of 0.05 the Official Baseline performance.