Benchmark-ready packaging of the SONAR synthetic-audio detection evaluation set
(arXiv 2410.04324), for speech anti-spoofing and
synthetic / deepfake voice detection.
SONAR is a binary-classification benchmark: bonafide (genuine human speech)
vs. spoof (AI-synthesized speech). The spoof side aggregates clips from eight modern
speech-synthesis systems, deliberately spanning architectures and providers so that a
detector cannot win by memorising one… See the full description on the dataset page:
https://huggingface.co/datasets/SpeechAntiSpoofingBenchmarks/SONAR.
