Thanks — you're asking about the average Time To First Token (TTFT) when using the Realtime API. Good question — TTFT is critical for natural, conversational voice experiences. Short answer • A practical target is under 500 ms TTFT. • For Advanced Voice integrations with LiveKit and GPT-4o, a common benchmark is around ~300 ms end-to-end TTFT, which is considered a human-like threshold for real-time speech interaction. What influences TTFT • Network latency between client and server (client upload and server response path). • Audio capture/encoding and packetization delays on the client. • Server-side model processing time (model size, prompt complexity, and realtime streaming behavior). • Any additional processing (speech-to-text, intent handling, or orchestration steps) before the model begins streaming tokens. What you can do to approach the 300–500 ms range • Use a low-latency network path and colocate services near your users. • Stream audio as soon as possible (small audio frames) rather than waiting for long buffers. • Choose model and inference settings tuned for low-latency streaming. • Minimize pre-processing or offload noncritical work so token streaming can begin quickly. If you’d like, tell me: • Where your clients are located (region) and whether you control the model/inference hosting, or • Which model and audio capture/streaming setup you plan to use— and I can suggest more specific tuning steps to help reach the 300–500 ms TTFT target. Sources: examples/bank-ivr | livekit/livekit_composite | DeepWiki | OpenAI and LiveKit partner to turn Advanced Voice into an API