Why perform speaker diarization at the end

[bofenghuang]

Hello @m-bain ,

Thank you for this excellent project!

From what I understand, the current pipeline merges the results of speaker diarization and STT at the end based on timestamps. I'm wondering why we don't just replace VAD with speaker diarization and pass the segments by speaker directly to Whisper (still need to ensure segments are <30s). Is it because we want to keep speaker diarization optional, or have benchmarks shown this approach performs better?