Turn-taking is the hard part of voice AI, and Alibaba’s Qwen team is shipping a stack built around it. Qwen-Audio-3.1 spans five models covering speech recognition, synthesis and realtime interaction. The centerpiece is full-duplex, deciding whether to keep listening, start speaking, stop or resume.
Three jobs get three models. One decides when to talk. Another writes the reply’s content. A third renders that text into streaming audio, conditioned on the conversation so far and on acoustic cues in the room. Training is split across stages the team labels Think, Act and Speak, and it leans on on-policy distillation from a text teacher rather than imitating prepared answers.
Gains and losses both go on the record. On Full-Duplex-Bench, the model cut its tendency to answer someone already engaged in another conversation, dropping from 0.13 to 0.03. Its filler rate fell from 0.759 to 0.296. Interruptions tell a worse story. The unwanted resume rate climbed from 0.035 to 0.130, and stop latency lands at 1.116 seconds, well behind the 0.383 recorded for GPT-Realtime-2, which also led a human red-team study.
Access is managed only. The realtime model runs on QwenCloud over WebSocket, with no open weights announced. Context is 262,000 tokens. Pricing is $6.40 per million audio input tokens, $0.80 per million text tokens and $24 per million output. A companion transcription model takes long offline audio.
