A Proficiency-Oriented Neural Network for Accent Classification

Abstract

Accents play a crucial role in how speech is perceived and interpreted, particularly in critical contexts such as corporate communication. Many accent classification models are built around nationality-based labels, which offer simple approximation of non-native speech patterns but limits in capturing the underlying linguistic features. To address this issue, we propose a new Proficiency-Oriented Neural Network for Accent Classification (ProNet-ACC) that classifies nonnative English accents using several public speech datasets. A key contribution of our approach is the conversion of nationality-based accent labels into a ranked language proficiency scale, providing a more informative and balanced framework for accent analysis. As part of an ongoing project, this work has developed an early-stage accent classification system with around 94% accuracy, establishing the foundation for future work on refining it with our own dataset and examining how accent and speech patterns shape audience perception and response in corporate communication.examining how accent and speech patterns shape audience perception and response in corporate communication.