I made my voice agent slower on purpose

Short summary

A voice AI sales-training builder found that an 880ms p95 pipeline was too fast — aggressive endpointing combined with low latency caused the AI to interrupt users mid-sentence. By relaxing end-of-turn thresholds (eager EOT off, timeout from 700ms to 1200ms), interruptions stopped with no user complaints about slowness. The key insight: latency and endpointing are coupled, and every millisecond shaved off the pipeline raises the cost of premature turn detection.

• •Fast pipeline + aggressive endpointing caused mid-sentence interruptions in a voice AI sales trainer
• •Relaxing end-of-turn thresholds eliminated cut-ins; nobody noticed the extra ~500ms delay
• •Latency and endpointing are not independent — speed amplifies endpointing errors; always quote p95 not p50

Generated with AI, which can make mistakes.