BUT+Omilia System Description VoxCeleb Speaker Recognition Challenge 2020
Authors:
Niko Brummer, Lukáš Burget, Ondrej Glembek, Pavel Matejka, Ladislav Mošner, Ondrej Novotný, Oldrich Plchot, Johan Rohdin, Anna Silnova, Themos Stafylakis, Shuai Wang, Hossein Zeinali, Omilia-Conversational Intelligence
Colaborators:
Brno University of Technology, Faculty of Information Technology, Speech@FIT, Czechia Omilia – Conversational Intelligence, Athens, Greece
Publication Date
06.11.2020
This is the system description corresponding to the systems developed by the BUT team for The Third DIHARD Speech Diarization Challenge. The systems for both tracks consist of a DOVERlap fusion of an end-to-end NN system with xvector based clustering systems in the form of spectral clustering and VBx. Given that the x-vector clustering systems do not provide overlapping speakers, overlapped speech is detected by a TasNet-based detector before the final fusion with the end-to-end approach.