Tag
This paper introduces CLD, a lightweight convex optimization-based language detection head for ASR that achieves 97-98% accuracy with under 100 training samples while reducing compute costs by 13x, addressing accent and dialect robustness across 5 languages and 24 sub-dialects.