VBHMM x-vectors Diarization (aka VBx)

Diarization recipe for winning system of track 1 of The Second DIHARD Diarization Challenge by Brno University of Technology. The recipe consists of

  • computing fbank features
  • computing x-vectors
  • doing Agglomerative Hierarchical Clustering on x-vectors as a first step to produce an initialization
  • apply Variational Bayes HMM over x-vectors to produce the diarization output
  • score the diarization output

Related publications:

DIEZ Sánchez Mireia, BURGET Lukáš, LANDINI Federico Nicolás and ČERNOCKÝ Jan. Analysis of Speaker Diarization based on Bayesian HMM with Eigenvoice Priors. IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH AND LANGUAGE PROCESSING, vol. 28, no. 1, pp. 355-368. ISSN 2329-9290. https://www.fit.vut.cz/research/publication/12139/

F. Landini, S. Wang, M. Diez, L. Burget, P. Matějka, K. Žmolíková, L. Mošner, A. Silnova, O. Plchot, O. Novotný, H. Zeinali, J. Rohdin: BUT System for the Second DIHARD Speech Diarization Challenge, accepted to ICASSP 2020, Barcelona

M. Diez, L. Burget, F. Landini, S. Wang, J. Černocký: Optimizing Bayesian HMM based x-vector clustering for the second DIHARD speech diarization challenge, accepted to ICASSP 2020, Barcelona

Link to GitHub:
https://github.com/BUTSpeechFIT/VBx