RADAR Challenge 2026 tests audio deepfake detection across 6 languages and real-world media distortions

The RADAR Challenge 2026, organized by researchers from multiple institutions, is an APSIPA Grand Challenge designed to push the boundaries of audio deepfake detection under realistic media conditions. Unlike typical benchmarks that use clean audio, this challenge simulates real-world distribution pipelines by applying compression, resampling, noise, and reverberation to audio samples. The dataset is divided into two phases: a development phase with labeled English utterances for analysis and paper writing, and a final multilingual evaluation phase containing over 100,000 utterances spanning six languages: English, Singapore English, Mandarin Chinese, Taiwanese Mandarin, Japanese, and Vietnamese. Systems are evaluated using the equal error rate (EER) metric for binary real/fake classification.

The challenge saw 33 teams submit systems during the development phase, with 22 advancing to the final evaluation. The reported results underscore that robust audio deepfake recognition remains a significant challenge, especially when dealing with multilingual data and varying media transformations. Many systems struggled to maintain low EER across all languages and distortion types. This suggests that current state-of-the-art models still lack generalization to real-world audio pipelines, highlighting the need for more diverse training data and adaptive detection techniques. The RADAR Challenge 2026 provides a critical benchmark for the community to improve deepfake detection systems, essential for combating misinformation and fraud in voice-based interactions.