• Does not need to rebuild the language model. A speech encoder turns audio into embeddings, an adapter projects them to the LLM’s embedding size, and those vectors replace an audio placeholder in the prompt.
  • Useful mental model: reshape speech until the LLM can attend to it like text. The encoder handles perception; the LLM handles the task.
  • Temporal subsampling matters because raw audio produces far too many frames for an LLM context window.

Connections