What is MOSS-Live?
MOSS-Live is a full-duplex voice model centered on turn-taking judgment. See its official benchmark figures, capabilities, architecture, and documented limits.
Updated ·
MOSS-Live: Full-Duplex Real-Time Voice Interaction Model
Announcement date:
MOSS-Live is a full-duplex real-time voice interaction model whose core capability is judging conversational timing: when to keep speaking, when to wait, and when to hand the turn back to the user. According to MOSS's officially published evaluation, it ranks first among the models shown in the chart on all six full-duplex metrics, including a 2.00% irrelevant-speech false-response rate and a 1.44% pause interruption rate. It is not uniformly ahead, however: multi-turn instruction following and audio understanding trail some competitors. This page updates the existing MOSS-Live profile with the newly published benchmark and capability details.
What MOSS-Live Is and What Full-Duplex Means
MOSS-Live is presented as a system whose core capability is judging real-time conversational interaction, including when to continue speaking, wait, or hand the turn back to the user. That turn-taking judgment is the organizing idea of the model rather than an add-on feature.
Full-duplex voice interaction means the system listens and speaks within the same continuous exchange instead of rigidly alternating between a listening state and a speaking state. In practice, the hard part is not generating speech but deciding whether incoming audio is meant for the model at all, and whether a pause is a finished turn or just a breath.
- Core capability: real-time interaction judgment, including continue, wait, or yield the turn
- Full-duplex framing: continuous perception and conversation rather than strict turn alternation
- Positioning: a voice interaction model, not only a speech-to-text or text-to-speech component
Official Benchmark Results on Six Full-Duplex Metrics
According to MOSS's official published evaluation, MOSS-Live ranks first among the models shown in the chart on all six full-duplex metrics. The most striking gaps appear on the two interruption-related measures.
MOSS-Live records a 2.00% irrelevant-speech false-response rate, compared with 68% for GPT-Live-1 and 75% for Doubao. It also records a 1.44% pause interruption rate, compared with 10.37% for GPT-Live-1 and 14.29% for Doubao. These figures describe how often the model responds to speech not directed at it, and how often it cuts in during a pause.
On tool use, the FDB v3 tool-call success rate reaches 51%, above GPT-Live-1's 45% and tied with Doubao for the highest. Post-interruption answer quality is 4.88/5 and entity tracking, correction and safety is 3.63/5, both the highest among the models in the chart.
- Irrelevant-speech false-response rate: MOSS-Live 2.00%, GPT-Live-1 68%, Doubao 75%
- Pause interruption rate: MOSS-Live 1.44%, GPT-Live-1 10.37%, Doubao 14.29%
- FDB v3 tool-call success rate: MOSS-Live 51%, GPT-Live-1 45%, tied with Doubao for highest
- Post-interruption answer quality: 4.88/5, highest in the chart
- Entity tracking, correction and safety: 3.63/5, highest in the chart
Where MOSS-Live Falls Short
The published evaluation is not a clean sweep. On τ-Voice multi-turn task success rate, MOSS-Live scores 57.93%, which is 40.55 percentage points above Doubao's 17.38% but still second among the six models in the original chart. GPT-Live-1 leads at 67.68%, a gap of 9.75 percentage points over MOSS-Live.
Audio understanding is another weaker area. Big-Bench Audio accuracy is 88.40%, below Doubao's 90.50%. Multi-turn instruction following is 3.35/5, below GPT-Live-1's 3.72/5 and Doubao's 3.44/5.
These results suggest a model that is unusually disciplined about when to speak, while still trailing on some sustained multi-turn and audio-comprehension tasks. The article states that the page does not disclose sample sizes, error ranges, API snapshots or the full evaluation configuration, so the figures should be read as vendor-published results rather than independently reproduced ones.
- τ-Voice multi-turn task success: 57.93%, second in the chart, 9.75 points behind GPT-Live-1
- Big-Bench Audio accuracy: 88.40%, below Doubao's 90.50%
- Multi-turn instruction following: 3.35/5, below both GPT-Live-1 and Doubao
- Undisclosed: sample sizes, error ranges, API snapshots, full evaluation configuration
How Real-Time Interaction Is Split from Background Reasoning
MOSS-Live divides real-time interaction from time-consuming tasks. The foreground handles continuous perception and conversation, while the background handles complex reasoning, tool calls and task execution. A coordination layer sits between them, sharing memory, context and task state.
This split matters for voice products because the two workloads have opposite requirements. Conversational timing demands low latency and constant attention; reasoning and tool calls take longer and should not block the conversation. Separating them lets the foreground keep the exchange responsive while the background works on the harder request.
- Foreground: continuous perception and conversation
- Background: complex reasoning, tool calls and task execution
- Coordination layer: shared memory, context and task state across both
Supported Capabilities
MOSS-Live supports full-duplex voice, semantic interruption, real-time video and screen sharing, and streaming simultaneous interpretation. Semantic interruption means the model can respond to the meaning of an interruption rather than only to a detected pause or volume change.
Real-time video and screen sharing extend the interaction beyond audio, while streaming simultaneous interpretation points to live translation use cases. The published material describes these as supported capabilities; it does not provide per-feature availability details, latency figures or regional rollout information.
- Full-duplex voice interaction
- Semantic interruption
- Real-time video and screen sharing
- Streaming simultaneous interpretation
Getting Started, Access and Pricing Status
The supplied material does not include a product URL, sign-up flow, API documentation, model snapshot names or pricing. The existing profile lists no website, and the published evaluation page does not disclose API snapshots or the full evaluation configuration.
For now, treat access and pricing as unknown rather than assuming a free tier or public API. If you are evaluating MOSS-Live for a product, the practical next step is to confirm availability, latency and pricing directly with the vendor before planning an integration.
- Access method: not stated in the supplied material
- API snapshots and documentation: not disclosed
- Pricing: unknown; no price information was supplied
- Suggested next step: confirm availability and terms with the vendor before integration planning
Profile Update Note
This page updates the existing MOSS-Live profile rather than creating a new entry. The previously maintained description reads: "MOSS-Live is a full-duplex voice model built around turn-taking judgment, with vendor benchmarks on interruption, tool calls, and multi-turn tasks." That description is preserved here as existing editorial content.
The new material adds the official six-metric benchmark figures, the capability list, the foreground/background architecture description and the documented gaps on multi-turn tasks and audio understanding. No search-demand signals were supplied with this update, so demand is treated as a new-term observation.
- Existing description preserved separately from this update
- New: official benchmark figures, capabilities, architecture, documented limits
- Demand status: new-term observation, no search signals supplied
Questions
What does full-duplex mean for MOSS-Live?
It means the model handles listening and speaking within one continuous exchange rather than strictly alternating states. MOSS-Live's core capability is judging when to continue speaking, wait, or hand the turn back to the user, which is what makes continuous interaction workable.
How does MOSS-Live compare with GPT-Live-1 and Doubao on interruption metrics?
According to MOSS's official published evaluation, MOSS-Live records a 2.00% irrelevant-speech false-response rate versus 68% for GPT-Live-1 and 75% for Doubao, and a 1.44% pause interruption rate versus 10.37% and 14.29% respectively. These are vendor-published figures without disclosed sample sizes or error ranges.
Is MOSS-Live the best model on every published metric?
No. It ranks first on all six full-duplex metrics in the chart, but on τ-Voice multi-turn task success it scores 57.93%, behind GPT-Live-1's 67.68%. Its Big-Bench Audio accuracy of 88.40% is below Doubao's 90.50%, and multi-turn instruction following at 3.35/5 trails both GPT-Live-1 and Doubao.
Does MOSS-Live support screen sharing and simultaneous interpretation?
Yes, per the published capability list: MOSS-Live supports full-duplex voice, semantic interruption, real-time video and screen sharing, and streaming simultaneous interpretation. The material does not provide per-feature availability, latency or pricing details.
Original sources
Tool Info
Has this tool changed? Help us keep it up to date
MOSS-Live Reviews
Loading reviews…