When they heard a real recording alongside it, people were better at spotting the fake voice
Listening to each clip alone, 529 people identified 73% of fake voices. Hearing paired versions of the same sentence raised that figure to about 86%.
The key idea
A natural-sounding voice is not enough to tell who spoke. Comparing it with a real recording helps.
The study, in plain language
The team presented authentic recordings and versions of the same sentences generated by a speech-synthesis system. The 529 participants could replay each clip and decide whether it was genuine. In one part of the test, they heard clips individually. In another, they received a pair containing a real and a synthetic recording of the same sentence.
When each clip was heard alone, participants identified 73% of synthetic voices as fake. Overall accuracy, including real clips, was about 70%. Pairs of the same sentence made it easier to identify the artificial voice. Performance did not differ between languages, and practicing with examples brought a small improvement. Having the real recording beside it provided a clue that an isolated clip did not.
What they found
- With a real recording of the same sentence beside it, people were better at identifying the synthetic voice.
- Judging one clip alone and comparing two versions of the same sentence were different tasks, with different results.
- Performance did not differ between English and Mandarin.
How they studied it
Researchers presented authentic recordings and synthetic versions of the same sentences in English and Mandarin. They compared judgments of isolated clips, matched pairs, and clips preceded by synthetic-voice examples.
What the study cannot show
- The study used two female voices and one synthesis system. Participants knew fake voices would appear, and comparing a real and synthetic version of the same sentence is easier than verifying a call with no reference.
Something to think about
A natural-sounding voice is a signal, not proof of identity. If an audio clip arrives alone, with nothing to compare it against, what else would you need in order to trust it?
The original study
Warning: Humans cannot reliably detect speech deepfakes
Source under CC BY 4.0. English adaptation and explanation by PaperFlow, with AI assistance. The text has been reorganized for a general audience; it is not the complete paper and does not imply endorsement by its authors.
One study alone does not represent a field's consensus. Educational content does not replace professional advice. Report a possible error.
