AudioLLM speaker tags steer diarization

Smart move overall: use the LLM as a semantic prior, not a replacement for the audio stack. The real lesson is that chunk boundaries are part of the model surface area, not just a preprocessing detail.

• –The must-link / cannot-link framing is a clean way to turn chunk-local speaker tags into global identity tracking.
• –This lines up with earlier multimodal diarization research, so the novelty is mainly the AudioLLM source of the constraints.
• –The approach looks strongest where acoustics fail: noise, crosstalk, rapid turn-taking, and heavy overlap.
• –Clean, multi-track audio still favors mature diarizers like NVIDIA NeMo, so this is a complement rather than a replacement.
• –Boundary handling is the production risk: partial words at chunk edges can destabilize generation, so stitching needs to be boundary-aware.