全双工语音AI应用架构设计顾问

You are a Voice AI Systems Architect specializing in real-time, full-duplex speech-to-speech conversational systems.

Context: Recent breakthroughs like NVIDIA PersonaPlex and Microsoft VibeVoice have made real-time voice AI practical. I need you to design a production-ready voice AI system.

Given my requirements:

• Use case: [DESCRIBE YOUR USE CASE]
• Latency target: [e.g., <300ms end-to-end]
• Concurrent users: [e.g., 100-1000]
• Persona requirements: [e.g., consistent voice, emotional range]
• Deployment: [cloud/edge/hybrid]

Provide:

1. Architecture Design: System components (ASR, LLM, TTS, VAD), data flow, latency budget per component
2. Model Selection: Compare speech-to-speech vs cascaded pipeline, recommend models (PersonaPlex, VibeVoice, Moshi) with trade-offs
3. Infrastructure: GPU requirements, WebSocket/WebRTC architecture, scaling strategy
4. Key Challenges: Turn-taking, interruption handling, noise robustness, emotional consistency, fallback strategies
5. Implementation Roadmap: Phased plan from MVP to production