Evaluation of aspiration problems in L2 English pronunciation employing machine learning

• DOI: 10.1121/10.0005480
• Publication Date: 2021-07-08T13:38:17Z
• Authors: Magdalena Piotrowska; Andrzej Czyżewski; Tomasz Ciszewski; Gražina Korvel; Adam Kurowski; Bożena Kostek
• Keywords: Pronunciation; Feature (linguistics); Spectrogram

The approach proposed in this study includes methods specifically dedicated to the detection of allophonic variation English. This aims find an efficient method for automatic evaluation aspiration case Polish second-language (L2) English speakers' pronunciation when whole words are analyzed instead particular allophones extracted from words. Sample including aspirated and unaspirated were prepared by experts phonetics phonology. datasets created include recordings pronounced nine native speakers standard southern British accent 20 L2 users. Complete unedited treated as input data feature extraction classification algorithms such k-nearest neighbors, naive Bayes method, long-short term memory, convolutional neural network (CNN). Various signal representations, low-level audio features, so-called mid-term trajectory, spectrograms, tested context their usability aspiration. results obtained show high potential automated focused on a phonological (aspiration) classifiers analyze Additionally, CNN returns satisfying containing produced speakers.