InterConv: Conversational Agent That Handles Interruptions Like Humans

Conversational agents built using large language models (LLMs) generally follow a turn-based interaction style, which differs from spoken human communication’s more dynamic nature. In natural conversations, people often interrupt each other during a dialogue for various reasons, such as seeking clarification, expressing disagreement, or responding to time constraints. Unlike LLM-based agents, human interactions are often less structured and rely on real-time responsiveness in spoken dialogue. Modifying LLM behavior to handle spoken interruptions typically requires significant resources that are not universally accessible. This paper presents InterConv, a straight-forward framework for creating an interruptible conversational Agent using off-the-shelf services. InterConv aims to enhance the naturalness of human-LLM voice interactions by enabling Agents to recognize and appropriately respond to user interruptions during speech, thus simulating a more human-like full-duplex conversation. The proposed method leverages existing speech-to-text (STT), response generation, and text-to-speech (TTS) technologies, making it accessible and easy to implement without specialized resources.