Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
University of California Los Angeles
4
0
2
AURA is proposed, an ultra-efficient representation-editing method that freezes the pretrained model and applies sparse scale-and-shift edits to decoder cross-attention heads, supporting dynamic activation editing as a practical path for grounding AED speech models.
Unified ASR for both children and adults is now achievable with a novel Entropy-Aware routing mechanism that enhances performance across diverse speech domains.
GC-LoRA reduces Word Error Rates by up to 10.9% in challenging acoustic environments while adding minimal trainable parameters.
Achieving state-of-the-art performance with just 10 hours of labeled data, Gumbel-BEARD revolutionizes self-supervised adaptation for speech models in low-resource settings.