The MRPL2 Non-native Speech Corpus

The MRPL2 Corpus is a non-native English speech dataset designed for use in mispronunciation detection and diagnosis (MDD), automatic speech recognition (ASR), and phonetic analysis. It contains 4,500 utterances from 23 non-native speakers from 14 native languages. Each utterance has been manually annotated to denote phonetic mispronunciations with respect to a British English dialect, Modern Received Pronunciation. The corpus includes audio files in both 48kHz and 16kHz, annotated TextGrids and word transcripts. Two master files are also included, which contain corpus-wide expected (exp) and ground truth (gt) phoneme sequences per utterance.