In the applications of computer assisted language learning, it is important to assess the pronunciation quality of second language learners in an automatic manner. Typically, this assessment is posed as a classification problem wherein the overall pronunciation quality is estimated at discrete levels. For classification, features are heuristically computed for an entire utterance considering factors influencing the pronunciation quality. However, the heuristic computation at the utterance level could not help in exploring the interdependencies between the factors and their effect at the sub-segment level. In this work, we learn the interdependencies between the factors by jointly modeling the labels representing the qualities of factors as well as pronunciation. Further, we also consider sub-segment level features for modeling. Experiments are conducted on data collected from Indian learners, considering the accuracy between the estimated qualities and the human expert ratings as performance measure. The highest improvements are found to be 19.13% and 14.93% (relative) when the proposed joint model is used with sub-segment and utterance level features respectively, and are compared to that of the baseline scheme without using a joint model.
Cite as: Srinivasan, A., Yarra, C., Ghosh, P.K. (2019) Automatic assessment of pronunciation and its dependent factors by exploring their interdependencies using DNN and LSTM. Proc. 8th ISCA Workshop on Speech and Language Technology in Education (SLaTE 2019), 30-34, doi: 10.21437/SLaTE.2019-7
@inproceedings{srinivasan19_slate,
author={Aparna Srinivasan and Chiranjeevi Yarra and Prasanta Kumar Ghosh},
title={{Automatic assessment of pronunciation and its dependent factors by exploring their interdependencies using DNN and LSTM}},
year=2019,
booktitle={Proc. 8th ISCA Workshop on Speech and Language Technology in Education (SLaTE 2019)},
pages={30--34},
doi={10.21437/SLaTE.2019-7}
}