On decoder-only architecture for speech-to-text and large language model integration