Developing an Automatic Pronunciation Scorer: Aligning Speech Evaluation Models and Applied Linguistics Constructs

Globalization and increases in the numbers of English language learners have led to a growing demand for English proficiency assessments of spoken language. In this paper, we describe the development of an automatic pronunciation scorer built on state‐of‐the‐art deep neural network models. The model is trained on a bespoke human‐rated dataset that reflects current perspectives on pronunciation and intelligibility. The new scorer is evaluated along three criteria: How well it explains expert human ratings, how it compares to other state‐of‐the‐art automatic pronunciation scorers in explaining expert human ratings, and the extent to which it exhibits bias toward different groups of test takers. Results indicate that the proposed scorer shows strong positive correlations with expert human ratings and outperforms other scorers. However, the scorer shows some bias related to audio quality and language family groups. We conclude with future directions for mitigating bias and argue that this scorer holds potential for use in operational settings.