Aurigin AI shows top-tier audio deepfake detection accuracy in new benchmark

Audio deepfakes have infiltrated call centers around the world with fraud attempts, but deepfake detection remains just as challenging for the voice channel as it is for other processes, like KYC checks with face biometrics. Aurigin AI cites the results of a new banckmark that shows its technology assures the legitimacy of voice biometrics with 96.75 percent accuracy.

Independent voice AI developer Podonos introduced the Audio Deepfake Detection benchmark to address what it considers a dangerous gap in the field.

Businesses struggle to assess the claims of audio deepfake detection providers in part because of a lack of useful data, according to Podomos. The company identifies the  ASVspoof 2019 LA as the closest thing to a public standard in the field. But ASVspoof 2019 LA predates the arrival of the advanced voice-cloning stack, and thus Podomos says, “Detectors trained on it do not generalize to current threats.”

Aurigin AI scored a 1.5 percent false positive rate (FPR) and a 5 percent false negative rate (FNR) in Podomos’ evaluation. Podomos assesses Aurigin as the most effective audio deepfake detection software for content moderation at scale, due to its low FPR.

Aurigin notes in a blog post that its infrastructure is about 100 times less costly than the only other deepfake detector in the same accuracy tier, at less than $0.001 per hour of audio analyzed.

A new leaderboard

Joining Aurigin at the top of Podomos’ leaderboard is Resemble AI, which scored 98.05 percent accuracy with a 1.4 percent FNR and a 2.5 percent FPR. Resemble’s model is assessed as the most effective for voice fraud screening and KYC due to its low FNR.

Podomos notes that Reality Defender has an unusual deepfake detection profile, with 71.3 percent accuracy, but a 53.7 percent FPR, and results returned more slowly than the other commercial models evaluated.

Hive’s audio deepfake detection model was found 83.5 percent accurate in the benchmark.

The other four audio deepfake detectors evaluated, which are all open-source models, accurately differentiated spoof attempts from real audio less than two-thirds of the time.

Article Topics

Audio Deepfake Detection  |  Aurigin.ai  |  biometric testing  |  deepfake detection  |  deepfakes  |  Hive  |  Reality Defender  |  Resemble AI  |  voice AI