Barge-In Is the Voice-Agent Feature Nobody Benchmarks. I Added It and Lost 120ms.

Short summary

Voice agent 'barge-in' functionality—letting users interrupt the agent gracefully—costs 120ms in latency increase due to false-positive guards, yet published benchmarks only report turn-taking latency and hide this critical tradeoff. The author reveals that interrupt-response latency and turn-taking latency pull in opposite directions, making them impossible to optimize simultaneously. Future benchmarks should measure minimum latency required to achieve accuracy targets, not report these metrics independently.

• •Implementing barge-in adds 120ms latency to handle false-positive interruptions at 95%+ accuracy
• •Published benchmarks measure only turn-taking latency, completely hiding the interruption-response tradeoff
• •Joint metrics (latency required to reach accuracy targets) should replace independent latency reporting

Generated with AI, which can make mistakes.