Self-Adaptive DNN for Improving Spoken Language Proficiency Assessment

Yao Qian, Xinhao Wang, Keelan Evanini, David Suendermann-Oeft


Automated assessment of language proficiency of a test taker’s spoken response regarding its content, vocabulary, grammar and context depends largely upon how well the input speech can be recognized. While state-of-the-art, deep neural net based acoustic models have significantly improved the recognition performance of native speaker’s speech, good recognition is still challenging when the input speech consists of non-native spontaneous utterances. In this paper, we investigate how to train a DNN based ASR with a fairly large non-native English corpus and make it self-adaptive to a test speaker and a new task, namely a simulated conversation, which is different from them monologic speech in the training data. Automated assessment of language proficiency is evaluated according to both task completion (TC) and pragmatic competence (PC) rubrics. Experimental results show that self-adaptive DNNs trained with i-vectors can reduce absolute word error rate by 11.7% and deliver more accurate recognized word sequences for language proficiency assessment. Also, the recognition accuracy gain translates into a gain of automatic assessment performance on the test data. The correlations between automated scoring and expert scoring could be increased by 0.07 (TC) and 0.15 (PC), respectively.


DOI: 10.21437/Interspeech.2016-291

Cite as

Qian, Y., Wang, X., Evanini, K., Suendermann-Oeft, D. (2016) Self-Adaptive DNN for Improving Spoken Language Proficiency Assessment. Proc. Interspeech 2016, 3122-3126.

Bibtex
@inproceedings{Qian+2016,
author={Yao Qian and Xinhao Wang and Keelan Evanini and David Suendermann-Oeft},
title={Self-Adaptive DNN for Improving Spoken Language Proficiency Assessment},
year=2016,
booktitle={Interspeech 2016},
doi={10.21437/Interspeech.2016-291},
url={http://dx.doi.org/10.21437/Interspeech.2016-291},
pages={3122--3126}
}