Robust speaker verification under noisy conditions

Determine how to achieve robust speaker verification when speech inputs are corrupted by background noise, ensuring reliable identity verification under noisy acoustic conditions.

Background

Speaker verification systems suffer substantial performance degradation in real-world environments due to diverse background noise such as babble, music, and non-stationary noise. Prior approaches include cascaded speech enhancement with verification and representation learning for noise-invariant embeddings, but each has drawbacks such as increased complexity or potential degradation of speaker discriminability.

This paper proposes a noise-conditioned mixture-of-experts framework to decompose the feature space into noise-aware subspaces and improve robustness. Despite the progress presented, the authors explicitly acknowledge that achieving robust speaker verification under noisy conditions remains unresolved, motivating further research in this direction.