New papers improve LLM-based audio-visual speech recognition

DoubleHelix introduces structured iterative cross-modal fusion for AVSR; another paper applies optimal transport-based semantic alignment to keep LLM-AVSR robust in adverse acoustic environments.