Speaker Verification

Speaker verification studies how to compare speech recordings to decide whether they belong to the same person, typically using learned speaker embeddings and similarity scoring. Systems built with x-vectors, ECAPA-TDNN, and self-supervised speech encoders are tested on benchmarks such as VoxCeleb and SITW, while anti-spoofing methods aim to reject replayed or synthesized voices. Current work also addresses cross-language and cross-domain robustness, fairness, uncertainty, interpretability, and efficient deployment, including evaluation with audio language models. A central challenge is that voices vary across sessions and conditions, so verification scores are probabilistic evidence rather than proof of identity; performance on standard benchmarks may not predict safety in real-world settings.