LiveKit provides a framework for streaming multiple AI model components—such as STT, LLM, TTS, and VAD—in a low-latency pipeline using the VoicePipelineAgent. This agent coordinates the flow between each component, allowing you to build real-time, multimodal applications. To reduce latency between these models, LiveKit offers several strategies and features: • Network Proximity: Deploy your agent close to your LLM, STT, and TTS providers to minimize network round-trip times. • Provider Selection: Choose low-latency providers for each pipeline component. Performance can vary significantly between providers. • Buffering: Use pre-connect buffering to start processing audio as soon as it arrives, reducing initial wait times. • Metrics Monitoring: Leverage the Agents metrics API to track and optimize latency at each stage (VAD, STT, LLM, TTS). Key metrics include Time To First Token (TTFT) and Time To First Byte (TTFB). LiveKit's architecture enables you to monitor and optimize each stage of the pipeline, and real-world deployments have achieved sub-second latencies by following these best practices. If you need implementation details or code examples for integrating these components with LiveKit, let me know which part of the pipeline you'd like to focus on. Sources: Voice AI quickstart | LiveKit Docs | An open source stack for real-time multimodal AI | AI Agents Framework | livekit/livekit_composite | DeepWiki | README.md | VoicePipelineAgent | LiveKit Docs