MOSS-Live: A Full-Duplex Real-Time Interaction System

Keeps listening as it speaks, distinguishing agreement, a thoughtful pause, and a real interruption

Complex tasks run in the background while the conversation continues, making real-time voice interaction feel more human

Try it now

API access

LunaStart a call

LeoStart a call

Listens like a person, gets things done like an assistant

Listens while speaking and uses meaning to know when to respond, wait, or stop

When needed, it can also see, translate, and get things done

A natural conversational rhythm

It keeps speaking when you signal agreement and stops immediately when you interrupt. If you pause to think, it waits for you to finish.

Knows who is addressing it

With several people present, it knows which remarks are addressed to it and stays out of side conversations.

Remembers where it left off

When interrupted, it knows exactly which word it reached and picks up from there, without starting over.

Translates as it listens

It translates into another language as you speak, without waiting for you to finish the whole passage.

Gets things done as you talk

Ask for live information and it looks it up. Research and planning run in the background, and it shares the results at a suitable moment.

Sees and takes initiative

It supports video calls and screen sharing. Agree on a rule, and it will speak up when the conditions are met.

Listen and speak together. Keep thinking without interrupting the conversation.

Full-duplex interaction lets listening and speaking happen together

Foreground–background coordination handles complex tasks in the background while your conversation continues

Listening and speaking happen together

Most voice assistants work like walkie-talkies: you finish, then they speak. MOSS-Live feels more like a phone call. It keeps receiving and understanding input while speaking, so it can interpret your voice at any moment.

You

I'd like to book a...ThinkingA ticket for next WednesdayHold on

MOSS-Live

Mm-hmmWaits for you to finishSure, next Wednesday...

Agreement keeps it goingPauses give you spaceInterruptions stop it immediately

Understands meaning, not just milliseconds of silence

Whether you have finished depends on what you said and what you still intend to say. When voices overlap, it first works out whether you are agreeing, interrupting, or talking to someone else before deciding how to respond.

Traditional voice assistants · A silence threshold ends your turn

You: I'd like to book a...PauseJumps in: What would you like to book?

MOSS-Live · Semantic turn detection

You: I'd like to book a...Pause · It waits...ticket for next WednesdaySure. Where to go next Wednesday?

“Mm-hmm” · “Right”

Recognizes agreement and keeps speaking

“Hold on”

Recognizes an interruption and stops immediately

Side conversation

Not addressed to it, so it stays quiet

Remembers only what it actually said

When interrupted, it knows exactly which word it actually reached. The conversation continues from that point, without assuming you heard anything it had not yet said.

1, 2, 3... 17, 18, 19

Already spoken · Included in context

20, 21, 22...

Not yet spoken · Excluded

The moment you say “stop”

You: What number did you get to?I reached 19. Shall I continue from 20?

Chats in the foreground, works in the background

The foreground interaction model stays available for real-time listening and speaking, while the background model handles complex reasoning and tool calls asynchronously. They share memory, context, and task status, bringing results back into the conversation at an appropriate moment.

Real time · Always available

Foreground interaction model

Memory · Context · Task status

Coordination layer

Asynchronous · Conversation keeps flowing

Background model

You

Research this for meTalk about something else

Foreground

SureRespondShare results

Background

Researching...Done