QwenAudio/qwen-audio-agent

A realtime voice runtime that keeps Agents talking, working, and present. Real-time Voice Runtime for AI Agents

What it solves

Qwen Audio Agent provides a realtime voice runtime that prevents the "waiting game" in AI conversations. It solves the problem of conversation gaps that occur when an agent needs to perform background tasks, call tools, or process complex information, ensuring the assistant remains present and interactive while work is being done in the background.

How it works

The project uses a three-layer architecture that separates the voice frontend from the backend agent. The voice frontend handles the immediate, full-duplex realtime interaction (listening and speaking), while the backend agent executes asynchronous tasks (tool use, file work, coding). This allows the user to continue chatting or ask for progress updates while the backend agent works on a task, with results naturally returning to the conversation once complete.

Who it’s for

It is designed for developers building voice-driven AI assistants for desktop productivity, smart cockpits, customer service, and potentially embodied intelligence, as well as users who want a persistent, low-latency voice interface for their existing AI agents.

Highlights

  • Full-duplex interaction: Supports natural interruption and sustained multi-turn voice conversations.
  • Modular Frontends: Compatible with various cloud services (Qwen, OpenAI, Gemini, Doubao, StepAudio) and local deployments (Hugging Face Speech-to-Speech, MiniCPM-o).
  • Agent Integration: One-click integration with a wide array of backend agents (e.g., Qwen Code, Claude Code, DeepSeek Harness) via the ACP architecture.
  • Multi-platform UI: Available as a WebUI, terminal TUI, and a desktop floating orb for macOS, Windows, and Linux.
  • Cros-session Memory: Includes long-term per-user personalization and memory.

Related

  • Project
    QwenLM/qwen-code

    An open-source AI coding agent framework that integrates into terminals, editors, and chat apps, supporting multiple LLM providers and local models.

  • Dispatch
    Introducing Qwen: A Comprehensive LLM and LMM Framework

    Qwen is a project towards AGI consisting of a series of multilingual large language models (LLMs) and large multimodal models (LMMs), including open-source versions ranging from 1.8B to 72B parameters.

  • Dispatch
    Qwen2-Audio Release Notes

    Qwen2-Audio is an audio-language model capable of voice chat and audio analysis across more than eight languages and dialects, surpassing previous state-of-the-art performance on multiple benchmarks.

  • Project
    zhayujie/CowAgent

    CowAgent is an open‑source, self‑hosted AI assistant that plans tasks, runs tools, keeps a personal knowledge base, and can operate as a team of cooperating agents. It supports many LLM providers, multimodal input, dozens of chat channels (Telegram, Slack, WeChat, etc.), a plug‑in skill system, and a three‑tier memory with nightly “Deep Dream” distillation. Install via a one‑line script or Docker, manage everything through a web console, and extend the system through the public Skill Hub or custom tool integrations.

  • Project
    Pinvou/pinvou-agent

    Pinvou Agent is a cross‑platform desktop AI‑assistant that lets you work, design, or code with a conversational model. It provides persistent workspaces, file attachments, a local knowledge base, persona cards, and an extensible plug‑in system (Skills, MCP servers, CLI tools). The UI is built with React + Vite, wrapped by Tauri 2, while the Rust‑based CodeWhale engine handles model calls and tool execution. It runs with any OpenAI‑compatible endpoint or a local vLLM server, and all user data stays under `~/.pinvou3/`. MIT‑licensed and actively maintained.