Playbooks 11 min read· 22 Aug 2026· By Playbooks
Voice assistants — the modern architecture
STT → LLM → TTS is the old shape. Here's what teams are shipping now.
Traditional voice pipelines add ~2s of latency across three model calls. Modern pipelines interleave streaming to hit sub-500ms.
Streaming everything#
Send audio chunks to STT as the user speaks. Kick off LLM completion on the partial transcript. Stream TTS from the first sentence out.
End-to-end latency under 500ms is achievable in 2026 without exotic hardware.