Summary
AWS solutions architects Chintan Agrawal and Daniel Wirjo tackle one of the least-discussed engineering challenges in voice AI: turn-taking. The talk centers on a 200-millisecond constraint — the natural human conversation switching speed — and why existing production systems, including Salesforce’s best published result of 755ms from March 2026, still fall well short of that target.
The presentation walks through a complete voice pipeline architecture (STT → LLM → TTS), identifying Voice Activity Detection (VAD) as the critical but often overlooked component that controls when the agent listens versus responds. Three escalating implementation levels are covered with live Python demos using PipeCat: Level 1 uses Silero VAD, a 300K-parameter model with convolutional and LSTM layers that detects silence locally; Level 2 delegates turn detection to the STT provider (smarter, but opaque); Level 3 combines Silero with a local smart turn model that analyzes prosody and intonation to distinguish real interruptions from filler sounds like “um” or background coughs.
The code across all three levels is nearly identical — only the configuration changes — making the comparison directly actionable. The talk also breaks down a real latency budget from PipeCat creator Kwindla Kramer’s production measurements (approximately 40ms for STT, 52ms for network/jitter), illustrating exactly where time is lost and where engineers have room to optimize. A practical guide for anyone building production voice agents with AWS or open-source tooling.
📺 Source: AI Engineer · Published July 20, 2026
🏷️ Format: Tutorial Demo
