Convex Low-resource Accent-Robust Language Detection in Speech Recognition

Researchers introduced Convex Language Detection, a lightweight plug-in for ASR that improves language ID across accents and dialects.

Built with convex optimization and ADMM in JAX, it achieved about 97–98% accuracy with fewer than 100 training samples.

Tests across five languages and 24 sub-dialects, including Whisper and MMS-1B, showed lower cross-lingual decoding errors.

The method also cut compute by roughly 13x, making accent-robust speech recognition cheaper and more data-efficient.