The Chinese University of Hong Kong, Shenzhen, in collaboration with Amphion Technology, has officially unveiled the DuplexAgent-RSI full-duplex voice agent framework. This innovative solution seamlessly merges real-time voice communication, intricate task execution, and the autonomous refinement of collaborative strategies into a single, cohesive platform. It orchestrates real-time dialogue, background task management, and teamwork coordination into a continuously optimized system. This setup empowers users to adapt their needs on the fly during interactions, while background tasks are processed concurrently without disrupting the flow of conversation.

The framework dissects the entire collaborative workflow into six distinct modules, each with well-defined roles, facilitating pinpoint accuracy in identifying and resolving collaboration bottlenecks through task logs. Moreover, it establishes a self-perpetuating feedback loop that encompasses automated testing, fault identification, precise fixes, and paired validation. This loop enables iterative enhancements without altering the underlying model parameters, relying solely on refining the rules, settings, and operational logic of the collaborative control layer.

Experimental results showcase notable advancements in critical performance indicators, including voice-based Q&A accuracy, tool utilization efficiency, and real-time response times. These improvements translate into substantial boosts in overall collaboration effectiveness following multiple cycles of self-evolution. Looking ahead, the research team envisions leveraging testing and interaction data to fuel agent reinforcement learning, thereby expanding the scope of self-evolution from the collaboration layer to encompass the model layer as well. This vision aims to create a more holistic, recursive improvement cycle for full-duplex voice agents, propelling them into a new era of intelligent, adaptive interaction.