兄弟，我的回应在哪里？通过本地语音活动检测减少每次语音 AI 交互 600 毫秒

Building a voice AI on OpenAI's Realtime API often results in slower response times than necessary, primarily due to inference bottlenecks but also additional latency. By instrumenting a production telephony voice pipeline, it was found that local voice activity detection (VAD) can significantly reduce response time, achieving an average reduction of 689 milliseconds per turn for substantive responses. The findings detail how latency was measured and present a clean methodology, underscoring the importance for developers working with conversational AI on the Realtime API.