Dev note
Faster call greetings
Near-complete rework of the opening greeting to cut silence before the bot first speaks.
- Type
- New feature
- Shipped
- Areas
- Call · Performance
Greeting is now generated on a fast, dedicated model decoupled from the main conversation LLM. Generation starts before the caller connects. A persistent LLM socket is warmed the moment call config arrives, eliminating cold-socket setup cost. Greeting is streamed to TTS sentence-by-sentence so first audio plays sooner. Main LLM is warmed during greeting playback so the first real reply is fast too. A first-audio latency budget bounds worst-case greeting delay.