MOSS-Live leads in voice turn-taking but trails GPT on complex tasks
The new MOSS-Live model excels at natural conversation flow like handling interruptions, though it still struggles with multi-turn reasoning compared to OpenAI.
The short version
- MOSS-Live ranked first in six full-duplex metrics, including interruption handling and turn-taking.
- The model responded to speech not directed at it only 2% of the time, compared to 68% for GPT-Live-1.
- It correctly ignored irrelevant interruptions 68% of the time, significantly beating rivals.
- MOSS-Live scored 57.93% on the τ-Voice multi-turn task, trailing GPT-Live-1's 67.68%.
- The system supports live video, screen sharing, and streaming interpretation via a WebSocket API.
Why it matters
For users of voice AI, these results mean MOSS-Live will feel more 'human' during fluid conversations because it won't constantly cut you off or react to background noise. However, for complex workflows involving multiple steps, GPT-Live-1 remains the more reliable choice for now.
A new standard for natural flow
MOSI Intelligence and Fudan University's OpenMOSS team recently released a detailed evaluation of MOSS-Live. This full-duplex real-time voice system is designed to handle simultaneous two-way communication, which is the technical term for a system where both parties can speak and listen at the same time. The evaluation compared MOSS-Live against major competitors like OpenAI's GPT-Live-1 and ByteDance's Doubao SeedDuplex.
The results show that MOSS-Live leads in several key metrics that define how natural a conversation feels. For example, it only responded to speech not directed at it 2% of the time. In contrast, GPT-Live-1 responded to such speech 68% of the time, and Doubao SeedDuplex responded 75% of the time. This suggests MOSS-Live is much better at filtering out background noise or side conversations.
The model also excels at knowing when to stay silent and when to speak. It cut in during a user's pause only 1.44% of the time, while its rivals cut in between 10.37% and 14.29% of the time. Furthermore, it held its course through brief acknowledgments like "mm-hmm" 98.17% of the time. These capabilities are essential for creating a voice assistant that doesn't feel like it is constantly fighting for the floor.
How the dual-model architecture works
To achieve this level of responsiveness, MOSS-Live utilizes a unique dual-model architecture. It pairs a foreground model with a background model to split the workload of a conversation. The foreground model is responsible for the immediate tasks of listening, managing the turn-taking logic, and generating the spoken output for the user.
Meanwhile, the background model handles the heavy lifting of reasoning, making tool calls, and executing longer tasks. Because the two models share memory, context, and task state through a coordination layer, the system can maintain a coherent conversation. This allows the AI to keep talking to the user while it performs a separate action, such as searching for a hotel in the background.
This architecture is what enables the system to support advanced features like live video and screen sharing. By separating the 'speaking' logic from the 'thinking' logic, the developers can optimize for speed without sacrificing the ability to perform complex operations. The coordination layer ensures that the user never feels a disconnect between what the AI is saying and what it is actually doing.
Performance gaps in complex tasks
While MOSS-Live wins on the
Availability and future outlook
The OpenMOSS team has made MOSS-Live accessible to the public through several different channels. Users can test the system in an online playground or access it via a WebSocket realtime API. The specific model ID for this release is moss-live-1.0. This availability allows developers to begin integrating the model into their own applications to test its specific strengths in turn-taking.
The development of MOSS-Live follows previous releases from MOSI Intelligence, which includes the MOSS-Image and MOVA models. By releasing these models, the team is positioning itself as a significant player in the multi-modal AI space. The focus on full-duplex communication suggests that the next phase of AI interaction will move away from
Transparency and test limitations
The detailed evaluation of MOSS-Live was released in an official blog post by the OpenMOSS team on November 13. While the post provides a comprehensive look at where the model leads and where it falls behind its rivals, there are important caveats regarding the data. All the figures presented in the report are vendor-published, meaning the developers provided the results themselves.
Notably, the blog post does not disclose specific details regarding the full test settings used during the evaluation. This includes a lack of information on sample sizes and uncertainty estimates for the various metrics. Because these underlying details are missing, it is difficult to independently verify the exact conditions under which MOSS-Live achieved its high scores in turn-taking and interruption handling.
Nuances in instruction following
When looking at specific interaction styles, MOSS-Live shows a mixed performance in following multi-turn instructions. It received a score of 3.35 out of 5 in this category, which is lower than the 3.72 score achieved by GPT-Live-1. However, the model showed significant strength in maintaining high answer quality after an interruption, where it led the competition with a score of 4.88.
The model also faces challenges in how quickly it yields the floor to the user. Its rate of yielding the floor within two seconds was 52.49%, which is significantly lower than the performance of the Qwen Audio models, which exceeded 93%. These results highlight that while MOSS-Live is excellent at maintaining conversation quality, it may not yet be as quick as some competitors at stepping back to let the human speak.
What to watch
- How MOSS-Live's latency in tool-use scenarios improves in future updates.
- Whether other open models can match MOSS-Live's 2% response rate to non-directed speech.
Try it
Check out the MOSS-Live online playground to test how it handles your interruptions compared to other voice assistants.
Sources and links
- Reportingpandaily.com: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by pandaily.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
