Using Large Language Models to Shape Social Robots' Speech

Social robots are making their way into our lives in different scenarios in which humans and robots need to communicate. In these scenarios, verbal communication is an essential element of human-robot interaction. However, in most cases, social robots' utterances are based on predefined texts, which can cause users to perceive the robots as repetitive and boring. Achieving natural and friendly communication is important for avoiding this scenario. To this end, we propose to apply state-of- the-art natural language generation models to provide our social robots with more diverse speech. In particular, we have implemented and evaluated two mechanisms: a paraphrasing module that transforms the robot's utterances while keeping their original meaning, and a module to generate speech about a certain topic that adapts the content of this speech to the robot's conversation partner. The results show that these models have great potential when applied to our social robots, but several limitations must be considered. These include the computational cost of the solutions presented, the latency that some of these models can introduce in the interaction, the use of proprietary models, or the lack of a subjective evaluation that complements the results of the tests conducted.