EEND-Vector-Clustering

EEND-vector-clustering (nttcslab-sp/EEND-vector-clustering, ICASSP 2021, Kinoshita/Delcroix/Tawara) is NTT's hybrid answer to diarization's central tension. Pure end-to-end neural diarization (EEND) handles overlapping speech beautifully but chokes on long recordings and a variable speaker count; clustering approaches scale to hours and any number of speakers but can't do overlaps. EEND-VC splits the recording into chunks, runs EEND inside each chunk to get overlap-aware local diarization plus speaker embeddings, then clusters the chunk-wise embeddings across the whole recording to stitch speaker identities together. The Interspeech 2021 follow-up showed it beating the strong EDA-EEND baseline on real CALLHOME conversational data (2–6 speakers, overlaps included), and later work folded in trainable clustering (deep-unfolded infinite GMM) to cut speaker-confusion errors further. The repo is a Kaldi+PyTorch training/inference recipe (mini_librispeech and CALLHOME recipes included) — research code, not a product: it needs CUDA, a Kaldi build, and patience, and notably ships no license file, so check with the authors before building on it commercially. Architecturally it's the direct ancestor of ideas now standard in pyannote's EEND-VC pipeline and in several production systems.

• Outperformed EDA-EEND on CALLHOME real conversational speech (author-reported, Interspeech 2021)\n- Follow-up work reports reduced speaker-confusion errors with deep-unfolded iGMM clustering (author-reported)\n- No independent benchmark replication found