Voice and NPCs
MOSS-Live: full-duplex voice that keeps talking while tools run
OpenMOSS has launched MOSS-Live, a real-time system that listens while it speaks, tells a pause or an "mm-hmm" from a real interruption, and runs tools in the background without halting the conversation.
On 7 November, the OpenMOSS team published MOSS-Live, a voice system built for conversations that do not take turns neatly. It listens while it is speaking, decides whether a sound is a request to take the floor, and hands slow jobs to a background model so the talk carries on.
Most voice assistants still work like a walkie-talkie: you finish, they answer. MOSS-Live is aimed at the way people actually speak, with corrections, half-finished thoughts and other people talking nearby.
What it can do
OpenMOSS lists six capabilities:
- Full-duplex conversation. It listens and speaks at the same time, and separates a thinking pause or a brief “mm-hmm” from a genuine interruption.
- Selective attention. It uses context to tell speech aimed at it from background voices and nearby chatter.
- Continuity. It follows corrections and topic changes, and can return to an explanation it was interrupted in.
- Live vision. Camera or screen-sharing input joins the same conversation.
- Background tools. Research and tool calls run behind the scenes while the foreground model stays free to talk.
- Streaming interpretation. It can translate while the speaker is still talking, and speak up on conditions the user sets, such as a reminder.
The design pairs a foreground model, which handles listening, turn-taking and speech, with a background model for reasoning and tool use. A coordination layer shares memory, context and task state between them, so a result can be brought back into the conversation when it fits.
The numbers
The figures below are OpenMOSS’s own, run against GPT Realtime 2.1, GPT-Live-1, Doubao SeedDuplex and two Qwen Audio 3.0 models.
- It reports the best score of the compared models on all six interaction metrics it highlights, such as avoiding premature takeovers and ignoring speech not meant for it.
- Tool use on FDB v3: 51.00% success, tied for best.
- Multi-turn tasks on τ-Voice: 57.93% success, second to GPT-Live-1 (67.68%).
- Quality of the reply after an interruption: 4.88 out of 5, the highest in the FDB v1 table.
- Response latency after a user interruption: 1.38 seconds, slower than GPT Realtime 2.1 (1.26 s) and GPT-Live-1 (1.04 s).
It does not win everywhere. On Big-Bench Audio it scores 88.40%, behind Doubao SeedDuplex (90.50%) and GPT Realtime 2.1 (90.10%).
Access and limits
You can try it in an online playground or build on it through a realtime API; both are linked from the launch post. The GitHub repository was created on launch day and holds the README, benchmark charts and demos. It lists no releases, and I found no licence file or downloadable weights there, so treat this as a hosted service for now. The demos are travel and meeting assistants, not characters, and all results come from the team’s own evaluation.
Our take
Interruption handling is one of the things that makes a voiced character feel like a person rather than a menu. A game NPC that keeps talking while it checks the world state, and does not blurt over a player’s “wait, no”, is closer to the characters we describe in intelligent NPCs. MOSS-Live is built for assistants, not for characters with personalities, so the useful part for game makers is the foreground and background split. Expect that pattern to turn up in NPC dialogue systems.