Thank you for reaching out with your question! Replacing the VAD and EOU models with a custom model can indeed affect total latency, and I’d love to walk you through how this works: • Custom EOU models usually add about ~50ms of inference delay. While this is a small addition, it does play a part in the overall pipeline latency. • Transcription delay, which is around ~400ms, happens independently of the EOU model and often has the biggest impact on total latency. • One great benefit of custom models is that they can reduce false turn endings, helping you shorten endpointing delays and making the system feel more responsive. • Keep in mind, system bottlenecks like CPU-bound inference or handling high concurrency can also increase delays. Even though a 25ms improvement in your model’s performance might not drastically cut down total latency because of the larger transcription delay, it can still make a meaningful difference by smoothing out turn-taking, reducing interruptions, and allowing for more confident endpointing. As a next step, it might be helpful to check if transcription delays from your STT provider are limiting your performance, and also ensure your infrastructure is ready to support the computational demands of the custom model. Sources: Logs, metrics, and telemetry | LiveKit Docs | Improving voice AI's turn detection with transformers | livekit-agents/livekit/agents/metrics/base.py | Improving voice AI's turn detection with transformers