qwen-audio-agent

qwen-audio-agent is an open-source realtime voice runtime designed to keep AI agents present, conversational, and active even while executing complex background tasks. Rather than freezing or forcing silence when calling tools or querying external sources, it maintains continuous full-duplex interaction and returns task updates naturally into the live audio stream.

1. For the Non-Technical Reader

Imagine speaking with a colleague on the phone while they pull up a complicated report. Instead of placing you on silent hold, they keep talking with you, answer quick side questions, and seamlessly mention, "I've got that report ready now." qwen-audio-agent brings that natural human workflow to voice AI, ensuring conversations remain continuous even while the agent is actively working behind the scenes.

2. For the Technical Reader

Architecturally, the project relies on an open Agent Context Protocol (ACP) infrastructure optimized for asynchronous execution and voice interaction. Key technical features include:

• Full-Duplex Architecture: Supports natural interruption, continuous multi-turn state tracking, local VAD (Voice Activity Detection), STT, LLM, and TTS integration, along with Qwen3.5-Omni Realtime frontend support.
• Parallel Execution & Tooling: Background tasks run asynchronously alongside the main speech loop, preserving existing agent model configs, custom tools, Skills, and MCP (Model Context Protocol) integration.
• Client & Gateway Support: Includes desktop builds (macOS/Linux) with isolated data directories, WebUI, and Terminal UI (TUI) featuring live progress cards and customizable pet animations.

3. Why It Matters

Traditional voice assistants suffer from rigid, turn-based stop-and-wait limitations whenever a backend tool call occurs. By decoupling the conversation layer from background task processing, this project advances the state of real-time multi-tasking agents while offering a self-hostable open-source alternative to proprietary streaming APIs.

4. The "Voice AI Space Lab" Idea

The Real-Time Live Studio Co-Host: Build a live radio or stream co-host that conducts interactive voice Q&A with listeners while continuously dispatching background agents to pull real-time data, perform live fact-checks, and report back findings naturally mid-stream.

Check out the implementation on the qwen-audio-agent GitHub Repository.