Explainable dialect identification in audio using knowledge distillation and SHAP with temporal window

Life Cycle Reliability and Safety Engineering · Springer Nature · 2026-11-14 · DOI: 10.1007/s10772-026-10306-4

Аннотация

Studies on dialect identification have increased, with majority leveraging simple to complex classifiers. However, these complex systems exhibit opaque characteristics, such that their decision-making process is not transparent to the end-users. This limits the adoption, usability and acceptability of these systems due to the black-box nature of their predictions. In this study, an explainable dialect identification framework through an enhanced logits-based knowledge distillation approach is proposed. The approach involves transferring knowledge from an ensemble of advanced teacher classifiers to a smaller student classifier. Audio-based features are extracted from the IViE and Kannada dialect datasets and used to train the teacher and student classifiers. A novel sliding SHAP with temporal window technique was explored to explain the predictions achieved by the student classifier and account for the variations in sequential data over time. From the experimental results, we observe that incorporating an explainable knowledge distillation approach on a student classifier offers a robust solution when compared to traditional approaches.

Читать у издателя →

Похожие статьи