In this study, phoneme based Hidden Markov Models (HMMs) were used to evaluate pronunciation. First their suitability for this task was determined and second, the effectiveness of feedback at a segmental level in a pronunciation learning experiment was investigated. The study is based on ten monosyllabic Dutch words, spoken by native and non-native speakers of Dutch. Pronunciation was evaluated by an expert listener as regards nativeness. Words spoken by natives and judged native by the expert listener were used to train phoneme based HMMs. In a test of these models, the words judged non-native achieved significantly lower scores then words judged native. The Equal Error Rates were low enough to assume the HMMs suitable for pronunciation evaluation. Forty non-native second language learners of Dutch participated in a training experiment. Half of the group was presented with pronunciation feedback on word level, the other half got feedback on segmental level. We expected that the last group would be able to improve their pronunciation more than the first group. Test results confirmed this hypothesis.
Cite as: Sevenster, B., Krom, G.d., Bloothooft, G. (1998) Evaluation and training of second-language learners' pronunciation using phoneme-based HMMs. Proc. ETRW on Speech Technology in Language Learning (STiLL), 91-94
@inproceedings{sevenster98_still,
author={Bob Sevenster and Guus de Krom and Gerrit Bloothooft},
title={{Evaluation and training of second-language learners' pronunciation using phoneme-based HMMs}},
year=1998,
booktitle={Proc. ETRW on Speech Technology in Language Learning (STiLL)},
pages={91--94}
}