End-to-end neural diarization with vector clustering, or EEND-VC, is a framework designed for long-form conversational diarization. It works by making local speaker decisions in short windows and then using speaker embeddings to link those decisions across the full conversation. The quality of those embeddings is therefore central to the system's overall accuracy.

A new paper on arXiv proposes learning speaker embeddings that are conditioned on speaker activity, rather than relying on generic embeddings. The authors use real conversational data to train these activity-conditioned WavLM embeddings, which are intended to better capture the dynamics of who is speaking when.

The key idea is that by explicitly modeling speaker activity during embedding learning, the resulting representations are more suitable for the linking step in EEND-VC. This differs from conventional approaches that may use embeddings trained on more isolated or artificial conditions. The paper's contribution is thus a training strategy and embedding design aimed at improving long-form diarization performance.