feat(transcribe): add speaker diarization support

[basnijholt]

Summary

• Add speaker diarization as a post-processing step for transcription using pyannote-audio
• Identifies and labels different speakers in the transcript (useful for meetings, interviews, multi-speaker audio)
• Works with any ASR provider (Wyoming, OpenAI, Gemini)
• New optional dependency: pip install agent-cli[diarization]

New CLI Options

Option	Description
--diarize/--no-diarize	Enable speaker diarization
--diarize-format	Output format: inline (default) or json
--hf-token	HuggingFace token for pyannote models (required)
--min-speakers	Minimum number of speakers hint
--max-speakers	Maximum number of speakers hint

Output Formats

Inline (default):

[SPEAKER_00]: Hello, how are you?
[SPEAKER_01]: I'm doing well, thanks!

JSON:

{
  "segments": [
    {"speaker": "SPEAKER_00", "start": 0.0, "end": 2.5, "text": "Hello, how are you?"}
  ]
}

Usage Examples

# Install diarization extra
pip install agent-cli[diarization]

# Basic diarization
agent-cli transcribe --diarize --hf-token YOUR_HF_TOKEN

# Diarize a meeting recording with known participants
agent-cli transcribe --from-file meeting.wav --diarize --min-speakers 2 --max-speakers 4 --hf-token YOUR_TOKEN

Test plan

• Unit tests for DiarizedSegment dataclass
• Unit tests for align_transcript_with_speakers function
• Unit tests for format_diarized_output (inline and JSON)
• Unit tests for SpeakerDiarizer class with mocked pyannote
• Updated existing transcribe recovery tests with new parameters
• All 513 tests passing
• Pre-commit hooks passing