What happened
OpenMOSS released MOSS‑Live, an open‑source full‑duplex real‑time multimodal interaction system that enables continuous speech, simultaneous listening and speaking, and decoupled background tool execution for dialogue research.
On November 11, 2026, the OpenMOSS research team—comprising Fudan University’s Natural Language Processing Laboratory, the Shanghai Innovation Institute, and MOSI Intelligence—announced the release of MOSS‑Live. The system is open‑source and designed for real‑time multimodal interaction, supporting continuous conversational speech and simultaneous listening and speaking.
MOSS‑Live introduces a dual‑timescale architecture that separates rapid foreground conversational processing from slower background tasks such as tool execution or analytical reasoning. This decoupling allows the foreground module to operate on a millisecond timescale, preserving low‑ speech synthesis and recognition, while background processes can run without blocking the conversation.
The platform supports robust interruption handling, enabling users to interject at any point without waiting for a turn‑taking cue. It also processes ambient audio and visual streams, making it suitable for research into more natural, spontaneous dialogue scenarios.
The release includes code, documentation, and example integrations, positioning MOSS‑Live as a research‑grade foundation for developers seeking to build full‑duplex voice agents. No pricing or commercial licensing details were provided; the system is freely available under an open‑source license.
Why it matters
MOSS‑Live’s full‑duplex architecture marks a shift from traditional half‑duplex spoken dialogue systems, allowing more natural, interruption‑tolerant conversations and supporting real‑time processing of ambient audio and visual streams. By separating foreground conversational turn‑taking from ‑heavy background computation, the system preserves low‑latency interaction while still enabling complex tool calls, potentially accelerating research into more fluid human‑AI dialogue and expanding open‑source resources for developers building voice agents.
Full‑duplex voice interaction more closely mirrors human conversation, where speakers can overlap and respond instantly. Existing spoken dialogue systems typically enforce a strict turn‑taking model, limiting responsiveness and making interruption handling cumbersome.
By decoupling foreground and background execution, MOSS‑Live addresses a key bottleneck in voice AI: the introduced when an agent must wait for external tool calls or complex reasoning before responding. This architecture could enable richer, more interactive applications such as real‑time translation, live assistance, and multimodal assistants that process video and audio concurrently.
The open‑source nature of MOSS‑Live provides the research community with a concrete reference implementation for full‑duplex dialogue, potentially accelerating experimentation and standardization in this emerging area.
If adopted widely, the system could influence the design of commercial voice assistants, prompting a shift toward more fluid interaction models and prompting vendors to reconsider trade‑offs in their architectures.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Adoption of MOSS‑Live by academic labs and startups, integration with existing voice assistants, and any benchmarks comparing its and interruption handling to half‑duplex baselines.
Early adopters: Academic labs and AI startups may integrate MOSS‑Live into their prototypes, offering early performance data and use‑case demonstrations.
Benchmarking: Comparative studies measuring , interruption handling, and overall user experience against traditional half‑duplex systems will clarify practical benefits.
Community contributions: Open‑source contributions could extend the platform’s capabilities, such as adding new language models, tool integrations, or multimodal sensors.
Commercial interest: Larger tech firms might evaluate MOSS‑Live’s architecture for incorporation into proprietary voice assistants, potentially leading to licensing or partnership announcements.