Local Prototype Reconstruction for Speech-to-LLM Bridge Pretraining

Local Prototype Reconstruction (LPR) improves speech-to-LLM interfaces by ensuring bridge embeddings maintain compatibility with frozen LLM input-embedding neighborhoods. This lightweight regularizer outperforms standard next-word prediction and sentence-level contrastive pretraining in capturing token-level lexical compatibility.