AMECxSV slashes cross-lingual speaker verification error by adapting to metadata

A new paper from Xin Wei and colleagues introduces AMECxSV, an adaptive backend that improves cross-lingual speaker verification by fusing trial scores with metadata. Fixed front-end scores vary in reliability when languages mismatch, durations differ, or score sources change. AMECxSV treats metadata—such as language match, utterance duration, and score source—as calibration context rather than speaker evidence. It produces calibrated target posteriors and optionally abstains from low-confidence predictions. On a development-derived held-out split, the score+metadata heads reduced EER from 3.15% to 2.42% for TidyVoice and from 0.64% to 0.43% for LI-MSV. A dual-score head achieved 0.43% full-coverage EER.

At 79% coverage, the abstention mechanism yielded a remarkable 0.03% EER on accepted trials—though this is not a full-coverage metric. Controlled experiments with matched score-only, metadata-permutation, and metadata-only ablations confirm that metadata serves as calibration context rather than additional speaker evidence. The authors limit claims to metadata-available scoring settings. This work addresses a critical gap: standard ASV systems degrade in cross-lingual conditions, and metadata-driven calibration offers a lightweight, principled fix. The approach is particularly relevant for global voice biometrics, multilingual call centers, and forensic speaker recognition.