Learning Speaker-Activity-Conditioned WavLM Speaker Embeddings with Real Conversational Data

Overview & Synopsis

End-to-end neural diarization with vector clustering (EEND-VC) is a strong framework for long-form conversational diarization, where speaker embeddings are used to link local speaker decisions across windows. Conventional EEND-VC systems extract these embeddings by cutting out regions assigned to each local speaker, preferably using only non-overlapped speech. This works well when enough clean target-speaker speech is available, but it discards useful speaker evidence in overlapped regions and becomes unreliable when a local speaker appears mostly in overlap. We propose a speaker-activity-conditioned embedding extractor that adapts a pretrained WavLM-MHFA speaker verification model to extract target-speaker embeddings directly from full conversational windows. Inspired by the recently proposed diarization-conditioned Whisper (DiCoW) model for target-speaker ASR, frame-level diarization-dependent transformations inject local speaker-activity information into intermediate WavLM representations while preserving the pretrained initialization. Experiments on multiple conversational diarization benchmarks show that fine-tuning with limited real meeting data reduces speaker-confusion errors and improves EEND-VC diarization performance. The results further indicate that real conversational data provides an effective alternative to large-scale simulated-mixture training for diarization-aware speaker embedding extraction.