Add multi-speaker persona modes (Host/Guest)

[marc-shade]

Summary

When diarization identifies multiple speakers, create separate personas for each that can be chatted with individually or together.

Why This Matters

• Major "wow factor" for interview/podcast content
• "Ask the guest" is a compelling use case
• Differentiates from basic transcription tools
• Enables panel discussions with AI versions of speakers

Acceptance Criteria

• Detect and separate speaker content
• Create individual persona for each speaker
• Allow naming speakers (Host, Guest, etc.)
• Chat with individual personas
• "Panel mode" - chat with all speakers together
• Speaker-specific knowledge boundaries

User Flow

1. Upload interview video
2. Diarization identifies Speaker 1 & Speaker 2
3. User names them: "Marc" (Host), "Jane" (Guest)
4. Choose chat mode:
   • 💬 Chat with Marc
   • 💬 Chat with Jane
   • 👥 Panel Discussion (both respond)

Example Interaction

```
User: What's your take on AI regulation?

[Panel Mode]
Marc (Host): From what I discussed in the interview,
I believe regulation should be... [15:23]

Jane (Guest): My perspective, as I mentioned, is that
we need balanced approaches... [16:45]
```

Technical Notes

• Requires speaker diarization (Add speaker diarization (who spoke when) #3) as prerequisite
• Separate vector stores per speaker
• Persona prompts generated per speaker's content
• Consider speaker interaction patterns

Priority

🟡 Medium - High effort but major differentiator

---

Generated from LLM Council product review