Hugo Van hamme

Missing sensory streams typically break audiovisual tracking, but routing audio, face, and mouth inputs through modality-agnostic query attention shatters previous active speaker detection records across five benchmarks by up to 5.1 mAP.

Contextual reasoning in speech recognition can be dramatically improved by grounding models in rich video metadata, leading to significant error reductions in rare terms and named entities.

Fusing multiple speech encoders can dramatically elevate ASR performance without incurring significant computational costs.

Phoneme-first prediction boosts LLM-based speech recognition accuracy by reducing acoustic confusion and enhancing pronunciation understanding.

Forgetting in ASR models can be drastically reduced by focusing adaptation on low-energy subspaces, leading to superior performance across tasks.

Location-aware gating of low-rank adaptations unlocks SOTA dialectal ASR with <10% parameter updates and interpretable geospatial patterns.