Tag
The paper proposes a frontend-backend architecture for enabling tool calls in full-duplex speech models, where a speech frontend delegates to a text-based LLM backend for tool invocation, maintaining low-latency and natural interaction.