Add speaker diarization (who spoke when)

[marc-shade]

Summary

Automatically detect and label different speakers in the transcript (Speaker 1, Speaker 2, or custom names).

Why This Matters

• Table stakes for interviews, podcasts, meetings
• Enables speaker-specific persona generation
• Required for professional transcription workflows
• Differentiator vs basic Whisper implementations

Acceptance Criteria

• Automatic speaker detection and labeling
• Ability to rename speakers (e.g., "Speaker 1" → "Marc")
• Visual distinction between speakers in transcript
• Speaker labels included in exports (SRT, VTT, etc.)
• Per-speaker statistics (talk time, word count)

Technical Options

1. pyannote-audio - Best accuracy, requires HuggingFace token
2. whisperx - Whisper + diarization integrated
3. Simple-diarizer - Lightweight alternative
4. Manual labeling - Fallback for users who want control

Implementation Notes

# Example with pyannote
from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization")
diarization = pipeline(audio_file)

Priority

🔴 Critical - High effort but essential for target market

---

Generated from LLM Council product review