Recent end-to-end approaches have shown promise in extending large language models
(LLMs) to speech inputs, but face limitations in directly assessing and optimizing alignment …