Using deep learning to classify English native pronunciation level from acoustic information

CLR Phonetics Lab, University of Aizu, Tsuruga, Ikki-machi, Aizuwakamatsu, Fukushima-ken, Japan

^* e-mail: m5221101@u-aizu.ac.jp
^** e-mail: wilson@u-aizu.ac.jp

Abstract

The main purpose of this research is to test the use of deep learning for automatically classifying an English learner’s pronunciation proficiency, a step in the construction of a system that supports second language learners. Our deep learning dataset consists of 28 speakers – ranging in proficiency from native to beginner non-native – reading the same 216-word English story. In the supervised deep learning training process, we first label the English proficiency level of the data, but this is a complicated task because there are a number of different ways to determine someone’s speech proficiency. In this research, we focus on three elements: foreign accent, speech fluency (as measured by total number of pauses, total length of pauses, and speed of speech) and pronunciation (as measured by speech intelligibility). We use Long Short-Term Memory (LSTM) layers for deep learning, train a computer on differently labeled data, test a computer on separate data, and present the results. Features used from audio data are calculated by Mel-Frequency Cepstrum Coefficients (MFCCs) and pitch. We try several combinations of parameters for deep learning to find out what settings are best for our database. We also try changing the labeling method, changing the length of each audio sample, and changing the method of cross-validation. As a result, we conclude that labeling by speech fluency instead of by speech intelligibility tends to get better deep learning test accuracy.