Simple language normalization wins TidyVoice 2026 speaker challenge
A simple embedding-space tweak outperforms complex models on 38 unseen languages
Cross-lingual speaker verification suffers when training and test languages differ. The TidyVoice 2026 Challenge tackles this with 3,666 speakers across 40 languages for training, 808 for development, and 2,200 evaluation speakers in 38 completely unseen languages—all without language labels at test time. Most teams turn to complex multilingual models or meta-learning, but a new paper from Nina Hosseini-Kivanani takes a simpler route.
The author starts from the official SimAM-ResNet34 baseline pretrained on VoxBlink2 and VoxCeleb2 and fine-tuned on TidyVoice. The key innovation is revisiting Nuisance Attribute Projection (NAP): a compact language subspace is estimated from cross-language same-speaker differences, then embeddings are projected onto its orthogonal complement. After cosine scoring with Adaptive Symmetric score normalization, development EER drops from 2.97% (plain cosine) or 2.70% (AS-Norm) to 2.18%. The final Codabench evaluation score of 8.40 proves that a simple back-end normalization can rival far more complex systems submitted to the challenge.
- NAP reduces EER from 2.97% to 2.18% on the TidyVoice 2026 development set
- System handles 40 training languages and 38 unseen evaluation languages without language labels
- Final Codabench evaluation score of 8.40, competitive with complex approaches
Why It Matters
Proves that simple embedding normalization can solve cross-lingual mismatch, enabling cheaper, more robust voice biometrics.