Track: Research Track |
| A Novel Dataset for Testing Anti-spoofing Models in a Telephony Environment |
| In the last few years, synthetic voices have become incredibly realistic and more difficult to distinguish from authentic, human voices. Although impressive, these advances raise security concerns, increasing the need for models that can discriminate between human and synthetic voices under real-world conditions. While previous work has proposed datasets and models that provide convincing results for high-quality recordings, very few studies have examined the efficacy of models under diverse conditions - both speaker and channel variations. Thus, it is unclear how well these models generalize to novel, less pristine channel conditions. In this paper, we present a novel dataset for testing the performance of such models under noisy conditions associated with the cellular telephone network. We improve upon previous methods by including a variety of synthesizers as well as languages. Finally, we demonstrate that a model trained on this dataset can achieve high accuracy on novel telephony data without any degradation in accuracy on non-telephonic audio. |
|
| Presentation Video |
| Presentation Notes |
|
HOUGHTON-ANovelDatasetforTestingAnti-spoofingModelsinaTelephonyEnvironment.pdf |
Real Time Communications Conference & Expo at Illinois Tech
IEEE International Conference
Follow Us
Share This
