With the advancement of real-time speech models, full-duplex has emerged as a critical technical direction.
This enables speech models to "listen and speak simultaneously", shifting the core question to: when listening and speaking happen at the same time, how should the system make continuous decisions to naturally participate in conversations?
Turn-taking detection, interruption recognition, listening feedback, rejection of non-directed speech, and the impact of new information on subsequent generation all need to be reworked within this framework. Evaluation benchmarks such as Full-Duplex-Bench are breaking down these behaviors into quantifiable dimensions.
Starting from these questions and combining real conversation scenarios, we will explain what true full-duplex real-time interaction is, and introduce MOSS-Live’s capabilities and evaluation performance in this field.
What is Full-Duplex Voice Interaction?
In voice interaction, the core distinction between half-duplex and full-duplex lies in whether input and output can occur concurrently.
A full-duplex system can listen and speak at the same time. While continuously processing incoming audio streams, the model generates outgoing audio streams, supporting on-the-fly interruption, feedback signals, overlapping speech processing and natural turn-taking.
By contrast, a half-duplex system can only either "listen" or "speak" at any given moment, relying on voice activity detection (VAD) for turn switching.
The full-duplex real-time interaction we focus on encompasses two dimensions: listening and speaking run in parallel in time, and conversational behaviors are continuously adjusted based on semantics and context.
The former allows the system to listen while speaking; the latter determines how it engages in communication.
The key criterion to judge whether a system possesses true full-duplex interaction capability: can information heard while the model is speaking promptly feed into subsequent conversational decision-making.
Simply keeping the microphone active, or recording user audio first and processing it only after the response finishes playing, does not prove the model is truly listening and speaking in parallel. On the other hand, full-duplex does not guarantee perfectly timed responses every time. Parallel listening and speaking is the foundational capability; how this capability is applied shapes the user interaction experience.
Why Interruptibility Alone Is Not Proof of Full-Duplex
"AI stops speaking as soon as you start talking" is often viewed as an intuitive feature of full-duplex, yet this same surface effect can be achieved via different technical mechanisms.
For example, a system can use VAD to detect user speech and pause playback once voice is detected, then process the user input afterwards. This mechanism enables interruption, but cannot independently prove that the model continuously understands new information during output.
A more challenging question: should the AI always stop when the user starts speaking?
Suppose the AI is explaining a proposal. When you say "Got it", you are signaling that you are following along. If you say "Wait, the budget is only 2000", you are modifying constraints. If both utterances trigger the same stop action, the conversation still feels rigid.
When the AI presents the revised proposal and you softly reply "Okay", such brief affirmations usually do not trigger turn switching. You do not need to stay silent for fear of interrupting it.
Similarly, when you describe your requirements at length, a capable system can provide short responses at appropriate moments to indicate it is still listening. This type of response is called backchannel, or listening feedback.
Naturally, continuously capturing ambient audio also means the system must judge which speech is addressed to it. When you turn to talk to a colleague or navigation audio plays nearby, the AI should not treat every human voice as a new instruction.
Full-Duplex-Bench v1.5 differentiates overlapping speech scenarios including user interruption, user affirmation, third-party conversation and background human voices, precisely because they require distinct handling strategies.
Therefore, interrupt capability serves as an entry point to observe full-duplex experience, but cannot be the sole benchmark. We also need to examine what the model hears, how it makes judgments, and whether subsequent replies truly incorporate newly added information.
MOSS-Live: A Full-Duplex Real-Time Interaction System for Continuous Communication
MOSS-Live is a full-duplex real-time interaction model built according to the above criteria.
It continuously understands new input while outputting speech. Information captured mid-utterance feeds into conversational decision-making in real time: thinking pauses are not misrecognized as turn endings, brief affirmations do not interrupt output, supplements and corrections can directly alter subsequent replies, and third-party conversations or background human voices are not mistaken for new instructions. On this basis, MOSS-Live integrates visual perception and tool calling, extending this interaction paradigm to richer application scenarios.
In public multi-person meeting demos, MOSS-Live can distinguish attendees’ brief affirmations from new questions and proactively join discussions after information retrieval. In a self-driving tour planning demo, MOSS-Live can resume unfinished topics after users repeatedly adjust requirements and switch subjects, while remaining immune to navigation broadcast audio. These scenarios demonstrate MOSS-Live’s interaction performance in real-world environments.
Full capability introductions, demo recordings and evaluation results are available on the MOSS-Live detail page.
Try MOSS-Live Now
MOSS-Live is now open for online trial and API integration.
You can chat with MOSS-Live on the MOSS Intelligence official website, or integrate the real-time interaction API into your own products and applications.
- Try MOSS-Live online: https://platform.moss.cn/app/playground/
- Real-time interaction API documentation: https://platform.moss.cn/docs/scenarios/moss-live/
- Project introduction and full evaluation: https://github.com/OpenMOSS/MOSS-Live