Revisiting Accent Classification with DistilHuBERT: Speaker-Disjoint Evaluation and Cross-Corpus Generalization
DistilHuBERT accent classifiers look strong in-corpus but fail speaker-disjoint and cross-corpus tests.
The study re-evaluates DistilHuBERT accent classification using L2-ARCTIC and several external English accent datasets. It compares file-level and speaker-disjoint splits, plus balanced external-data training and teacher-student pseudo-labeling. The large gap between in-corpus and external results shows file-level evaluation overstates robustness, while speaker-disjoint testing better estimates generalization to unseen speakers. Models trained only on external data did not reliably generalize to L2-ARCTIC, and confidence filtering alone was insufficient under domain mismatch.
- File-level splits overestimate accent-classification robustness.
- Speaker-disjoint evaluation better estimates unseen-speaker generalization.
- External-only training does not reliably transfer to L2-ARCTIC.
- Confidence-based pseudo-label filtering fails under domain mismatch.
- AI models
- DistilHuBERT
From the original article
1 min read at arxiv.orgRecent accent classification systems can achieve high accuracy in standard in-corpus evaluations. However, when speakers overlap between the training and test sets, such results may overestimate model robustness compared with real-world deployment scenarios. This study revisits the robustness of DistilHuBERT-based accent classification by using L2-ARCTIC and multiple external English accent datasets.
By Xu He · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.