Qwen's realtime API adds speech-to-speech for voice agents
The plus/flash models expose an OpenAI-Realtime-style WebSocket with barge-in and tool calls — and Qwen's TTS sibling now tops the Speech Arena.
Copy markdown
Two models: plus for quality, flash for latency
Qwen-Audio-3.0-Realtime went live on Alibaba's Model Studio API on Aug 10 in plus and flash variants — full-duplex speech-to-speech that hears you and talks back over a single WebSocket. Flash chases low latency; plus chases output quality.
The event API mirrors OpenAI's Realtime schema
It speaks the same vocabulary — session.update, input_audio_buffer.append, response.audio.delta, response.done — with 16 kHz PCM in and 24 kHz out over WebSocket or WebRTC. If you already built on OpenAI's Realtime API, most of your client carries over.
Barge-in, smart-turn, and tool calls are built in
The model cancels its own response the moment you interrupt, offers server-VAD, semantic smart-turn, and push-to-talk turn-taking, and supports function calling so it can hit your MCP servers and APIs mid-conversation. Tone and pacing shift with context.
The catch: Beijing-only, 50 turns, 5 minutes of context
Availability is currently the China (Beijing) region, and a session caps at 50 audio turns and 300 seconds of cumulative audio — plan resets for longer calls. Pricing lives in the Bailian console, not the public model page.
Elsewhere: Qwen's TTS sibling tops the Speech Arena
The matching Qwen-Audio-3.0-TTS ranks #1 on Artificial Analysis' Speech Arena — ahead of ElevenLabs v3 and Gemini 3.1 TTS — with ~300 ms first-packet latency, 16 languages, and 20 Chinese dialects. Alibaba now covers the full voice stack.