Researchers introduce a bilingual same-speaker evaluation set for five Iberian languages to analyze cross-lingual speaker verification under constant speaker identity, addressing the confounding of language mismatch with inter-speaker variability in standard protocols.
Applying this setup to a HuBERT-based system and analyzing results via the Cross-Lingual Transfer Matrix reveals that while speaker-related variability contributes to performance degradation, language mismatch remains the primary driver of cross-lingual loss.
These findings provide a more precise characterization of language dependence in cross-lingual speaker verification systems.