ISCA Archive IberSPEECH 2024
ISCA Archive IberSPEECH 2024

AUDIAS-UAM System Description for the Albayzin-RTVE 2024 Speaker Diarization Challenge

Alicia Lozano-Diez, Juan Ignacio Alvarez-Trejos, Laura Herrera, Beltran Labrador, Jeremie Touati, Sara Barahona

In this paper, we describe the speaker diarization system submitted by the AUDIAS-UAM team for the Albayzin-RTVE 2024 Speaker Diarization Challenge. Our primary submission consists of the combination via DOVER-Lap of three speaker diarization systems within the state-of-the-art: Pyannote, VBx and DiaPer. Both Pyannote and DiaPer systems are based on neural networks for diarization of shorter segments, followed by a matching algorithm to assigned predicted speaker labels for the whole recording. VBx is used to obtained speaker diarization labels over the whole recordings. The combination of these individual systems yields a 9.26% DER on our development set, with respect to a 12.26% DER of Pyannote, 15.30% DER of VBx and 23.25% DER of DiaPer, showing the potential of a fusion of three quite distinct diarization systems.