How acoustic voice deepfake detection works
An acoustic deepfake detector uses patterns learned from recordings of real and generated speech to estimate whether an input resembles synthetic audio. Microphones, codecs and background noise also change those patterns. A detector score cannot authenticate a speaker.
Updated
Listen to the recording
These files contain the same human performance. Listen to the consonants in “jacket” before and after telephone encoding. A change in timbre can have a mundane explanation.
“Don’t forget a jacket.”
“Don’t forget a jacket.”
Acted audio from CREMA-D. The recording catalog identifies the source files. Database license and contents license.
Separate the source from the channel
A classifier may learn features correlated with synthetic excitation, spectral texture or timing. It can also learn accidental cues, such as compression common to just one class in its training data. An unfamiliar microphone can change a real recording. A high-quality generator can leave few familiar traces.
The second file here was downsampled, μ-law encoded and decoded to PCM. Neither file was generated by a voice model. They are a control pair for channel sensitivity, not a detector accuracy test. No OrukZero inference result is asserted for this comparison.
Try the comparison
Start with consented human recordings and recordings of known generator provenance. Match codec and sampling conditions in both groups. Hold out speakers, generator families and channels, then report false positives on human speech and missed synthetic examples at the threshold you intend to use. Review provenance separately before acting on any individual score.
Inspect the OrukZero file API and score interpretation
Try it on your own audio.
The demo needs no account. The trial needs a card.