Streaming conversation research

A voice agent that listens.

Talk naturally, add a quick “mm-hmm,” or interrupt. Voice Light responds while keeping track of who has the floor.

Press Start microphone to wake the server.

For research and quality analysis, this demo privately stores microphone audio, transcripts, assistant responses, tool activity, and timing data for up to 30 days. Do not share sensitive information.

Current session

Conversation history

You waiting Assistant waiting

Your conversation will appear here as you speak.

Ephemeral model evidence

How the agent reads the floor

Last 20 seconds · dots are model observations, not conversation labels

Latest policy snapshot

These probabilities are causal evidence used by the hybrid policy. “User yielded” means the user is not holding the floor; “taking the floor” means speech should interrupt the assistant; “backchannel” means brief feedback that should let it continue.

Adapter
waiting
Silero speech
waiting
User yielded
—
Turn complete
—
Taking the floor
—
Backchannel
—
Final decision
waiting
Decision latency
—
First usable prediction
—
Latest sample
—
Playback action
—
Evidence cadence
no model observations
Session events