Real-time Voice Transport — Provider-Neutral Protocol
Public concept and protocol design for One's low-latency, full-duplex realtime voice transport, including client/server requirements, provider-neutral voice transport contract, and animated waveform visual feedback.
TL;DR: Design and requirements for One's low-latency, full-duplex realtime voice transport protocol, including a provider-neutral voice transport contract, smooth animated waveform feedback, and instant client-interruption handling.
Status as of 2026-07-01: see body.
Relations
The One platform communicates with users via highly conversational, responsive, and natural interfaces. Real-time Voice Transport establishes a standard, low-latency protocol and visual system that enables full-duplex conversational voice interactions across a wide variety of hardware devices and AI models.
By isolating the audio capture, transport serialization, and visual synthesis layers, the voice runtime ensures a continuous stream of user intent reaches the reasoning models while providing smooth, immersive feedback.
The Three Core Pillars
- Provider-Neutral Voice Transport Contract: Enforces a standard interface for real-time audio chunk chunking, synchronization, interrupt-handling, and stream state. Reasoning runtimes must conform to this contract rather than coupling to provider-specific WebSockets (e.g., Gemini Live API vs. OpenAI Realtime).
- Smooth Visual Waveform Feedback: The user interface displays a continuous, physics-modeled animated waveform embedded inside the Agent voice bar. This waveform acts as the heartbeat of the connection, reacting instantly to incoming audio metadata, decibel levels, and connection state.
- Full-Duplex Interruption Model: Implements client-side and server-side fast interruption detection. When the user speaks while the model is responding, the transport layer sends a deterministic cancel frame, immediately cutting the outgoing audio stream and resetting the model's response pipeline without lag or overlap.
Protocol Mechanics
The voice transport operates over low-overhead WebSockets or WebRTC data channels:
- Upstream (Client-to-Agent): Audio is chunked (PCM 16-bit, 24kHz or 16kHz mono) and packed into binary frames carrying a millisecond timestamp and stream ID.
- Downstream (Agent-to-Client): The server-side multiplexer forwards generated audio chunks along with volume envelope metadata (amplitude and frequency buckets) to drive the visual waveform on the client device.
- Interruption Signal: A high-priority control message (
"interrupt") with the active stream's target timestamp ensures both client and server immediately flush their playback buffers.