VBx

VBx (BUTSpeechFIT/VBx) is the diarization recipe from Brno University of Technology's Speech@FIT lab that became a fixture of the field's leaderboards. The pipeline is three stages: extract x-vectors from short segments, get a rough grouping with agglomerative hierarchical clustering, then refine it with a variational-Bayes hidden Markov model whose speaker-turn dynamics are modeled explicitly and whose emission distributions live in PLDA space. The Bayesian machinery matters because it lets the model decide how many speakers there are instead of being told — the Fb hyperparameter is the dial that penalizes creating new speakers. The repo's README publishes results with unusual protocol honesty (Forgiving/Fair/Full): CALLHOME at 4.42% DER forgiving (14.21% fair, 21.77% full), DIHARD II full at ~18.2–18.6%, AMI beamformed forgiving down to 2.8–3.9% — all with oracle VAD, so read them as clustering-quality numbers rather than end-to-end ones. The recipe ships with trained x-vector extractors so it runs standalone, and there's a random-initialization mode for files over 30 minutes where AHC gets slow. The line continues in DVBx (discriminatively trained VBx, ICASSP 2024) and DiariZen, the lab's newer toolkit. If you want a principled, inspectable, Apache 2.0 diarizer whose every design decision is documented in a 2022 Computer Speech & Language paper, this is it.

• CALLHOME: 4.42% DER (Forgiving), 14.21% (Fair), 21.77% (Full) — oracle VAD (author-reported, repo README)\n- DIHARD II Full: 18.19% dev / 18.55% eval; AMI beamformed Forgiving: 2.80% dev / 3.90% eval (author-reported)\n- No independent replication found; protocol reporting is unusually transparent