Innovative AI Approaches in Oral Proficiency Testing: Ethical Implications of ASR Systems

Oral proficiency testing plays a critical role in language assessment; however, classic ASR faces such problems as Americans Bias, which means difficulty for ASR as application to recognize non-American accents, and the evaluation inaccuracy. The major problem here is that classic ASR systems cannot recognize and assess non-American accents as they would reward non-biased, unbiased, and fair scores for speakers with regional or non-native English accents. This bias in particular distorts the fairness of the oral proficiency test. The current approaches are inefficient with regards to the fairness and standardization in natural language scenarios. Unfortunately, in practice the conventional ASR models are not perfect and become a source of errors and biases when trying to assign a score for the given speech; the specific problem with the current approach is that the speakers with regional or non-native accent will be almost impossible to score correctly. To address this bias, this study proposes BERTAdv, incorporating BERT which focuses on contextual understanding into the Adversarial training method for enhancing the ASR's accuracy in OPI. In an attempt to improve the BERTAdv to analyze both simple and complex linguistic attributes, the model is trained on a large and balanced dataset. The model presents desirable accuracy with a high R-squared of 0.93, MAE of 0.26, and RMSE of 0.46 and is more accurate and fairer than existing methods.