Alibaba's Qwen team has added Qwen-Audio-3.1-Realtime to its Qwen-Audio-3.1 family, a voice model built for full-duplex conversation rather than the usual record-then-respond pattern. In a full-duplex setup, audio flows both ways at once: the model listens while it is speaking, and it decides for itself when to jump in, stay quiet, or wait for the other side to finish.
That decision-making is the point of the release. Rather than relying on a separate voice-activity detector to mark when a user has stopped talking, the model is trained to judge turn-taking itself, reading pauses, interruptions and intent the way a person on a call does. Alibaba Cloud's documentation describes the model as converting real-time audio to text and speech over a streaming connection, with both input and output handled continuously rather than in batches.
Alongside the conversational core, the model supports tool calling and web search mid-conversation, so it can pull in live information without breaking the flow of speech. It is part of a five-model Qwen-Audio-3.1 release rather than a standalone product, sitting alongside other variants aimed at different latency and quality trade-offs.
This matters because full-duplex, low-latency voice has been one of the harder problems in conversational AI. Most voice assistants still wait for a clear pause before responding, which makes them feel stilted in natural back-and-forth. A model trained specifically to decide when to speak, rather than just when to transcribe, is a step towards voice agents that behave more like a real conversation partner than a dictation tool. Access is through Alibaba Cloud's Model Studio platform.
Source: MarkTechPost
