Summary

Reduces short speaker label jumps in the final diarization timeline.

• Add centroid-scored Viterbi smoothing inside each VAD segment
• Penalize short speaker-label switches while preserving sustained original label runs
• Collapse brief A-B-A label islands after temporal decoding
• Pass embeddings into timeline assembly so smoothing can use speaker centroid confidence
• Add regression tests for flicker removal and real speaker-turn preservation
• Refresh README/docs benchmark numbers and explain temporal smoothing

Results

VoxConverse dev benchmark:

Validation

• python -m ruff check .
• python -m ruff format --check .
• python -m pytest tests/test_diarize.py --cov=src/diarize --cov-report=term-missing -q
• python -m mkdocs build --strict --site-dir /private/tmp/diarize-site