Full-Duplex Voice: What Changes When the Model Can Be Interrupted

Alibaba Qwen released Qwen-Audio-3.1-Realtime, a full-duplex voice model capable of reasoning, calling tools, and deciding when to speak. This technology enables simultaneous communication, replacing turn-based interactions with immediate barge-in capabilities. On τ-Voice adaptation, task success increased from 78.4% to 82.0%, while responses to background speech decreased from 73% to 13%. The model requires separate playback and microphone streams to ensure immediate interruption and proper echo cancellation. Qwen-Audio-3.1-Realtime is currently available via API on QwenCloud.