Embrace the Urgency: Short but Precise Agent Speech via LLM and Speech Generation

In a time-pressured kitchen, a collaborative agent that talks too much slows both itself and its human partner down. This thesis explores how such an agent can keep its spoken responses short yet precise. Using a language model to draft utterances and speech synthesis to deliver them, the project investigates how to constrain, time, and ground the agent’s speech so that it stays brief, task-relevant, and well-timed instead of verbose. The goal is a concise communication policy that preserves helpfulness while minimizing how long the agent speaks.