X-AuT reduces speech LLM inference costs by progressively pruning audio-encoder layers through representation alignment and cross-scale distillation. The framework uses LoRA adapters and scheduled student-policy supervision to prevent deletion and end-of-sequence errors while keeping the language model backbone frozen.
HOW THIS AFFECTS YOU
●
builderYou can lower inference latency and costs for speech-based AI products.
●
researcherThis method provides a way to prune audio encoders without disrupting decoder embeddings.