OpenAI DevDay 2026GPT-Live-1 APISub-150ms LoopWebRTC Data ChannelVoice Tool Execution
OpenAI GPT-Live-1 Voice API & Native Speech Engine Studio
Deploy direct audio-in to audio-out real-time voice streaming pipelines. Eliminate cascaded speech-to-text latency using GPT-Live-1's native multimodal voice weights, instant acoustic barge-in interruptions, and live asynchronous tool calling.
Audio-to-Audio Latency
< 140ms P95
Direct neural audio loop
Streaming Transport
WebRTC Opus 24kHz
Full-duplex peer connection
Barge-in Interruption
Instant < 20ms
Acoustic echo cancellation
Voice Tool Execution
Async Parallel
Zero audio playback stutter
🏗️ Sub-150ms Native Voice & WebRTC Architecture
GPT-Live-1 Engine
💻 Infrastructure & Software Technology Stack
🎙️ Real-Time Audio Infrastructure
- Transport: WebRTC MediaStream & Full-Duplex WebSocket
- Codec: Opus Audio 24kHz @ 32kbps low-latency chunking
- VAD Engine: Server-Side Voice Activity Detection with 200ms silence threshold
- Jitter Buffer: Dynamic packet replay & acoustic echo cancellation
🧠 Foundation AI Engine
- Model: OpenAI GPT-Live-1 / GPT-4o Realtime Audio Weights
- Latency: Direct cross-modal audio-to-audio (~80ms TTFT)
- Tool Calling: Asynchronous function dispatch during live speech
- Barge-In: Instant audio stream truncation on user interruption
☸️ Gateway & Orchestration
- Deployment: Kubernetes Service & Ingress with sticky sessions
- Gateway Runtime: Python 3.11+ / FastAPI / WebSockets
- Client SDK: Vanilla JavaScript WebRTC PeerConnection
- Observability: Prometheus audio packet loss & round-trip latency HUD
⚙️ Audio Streaming & Voice Configurations
💡 Human Experience & Zero Latency Impact
By shifting from sequential Whisper-to-LLM-to-TTS cascades to GPT-Live-1's direct neural voice-to-voice stream, end-to-end conversation turnaround drops from 2,800ms to sub-140ms, achieving true human conversational parity.
Loading GPT-Live-1 Voice API configuration...