Experiments on Non-native Speech Assessment and its Consistency
Ziwei Zhou, Sowmya Vajjala, Seyed Vahid Mirnezami · NPARC · 2019
In this paper, we report some preliminary experiments on automated scoring of nonnative English speech and the prompt specific nature of the constructed models. We use ICNALE, a publicly available corpus of non-native speech, as well as a variety of non-proprietary speech and natural language processing (NLP) tools. Our results show that while the best performing model achieves an accuracy of 73% for a 4-way classification task, this performance does not transfer to a cross-prompt evaluation scenario. Our feature selection experiments show that most predictive features are related to the vocabulary aspects of speaking proficiency.