Towards Quantifying and Reducing Language Mismatch Effects in Cross-Lingual Speech Anti-Spoofing

Problem Statement

• Language mismatch negatively impacts the performance of speech anti-spoofing systems.
• Existing anti-spoofing datasets are predominantly in English. Acquiring multilingual datasets, especially for low-resource languages, is costly and challenging.
• There is a need to quantify the extent to which language mismatch affects anti-spoofing systems.

Take Home Message

• It is estimated that language mismatch effect can cause a relative performance reduction of over 15%. This highlights the need for models that are robust across different languages.
• Data augmentation using TTS with diverse accents (ACCENT) can effectively mitigate language mismatch effects
• The ACCENT method is promising for multilingual and low-resource language scenarios.

BibTex

@article{QuantifyingLanguageMismatch2024,
  title={Towards Quantifying and Reducing Language Mismatch Effects in Cross-Lingual Speech Anti-Spoofing}, 
  author={Tianchi Liu and Ivan Kukanov and Zihan Pan and Qiongqiong Wang and Hardik B. Sailor and Kong Aik Lee},
  year={2024},
  eprint={2409.08346},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  url={https://arxiv.org/abs/2409.08346}
}