Humans unable to detect a quarter of deepfake speech, study finds
Research conducted by the University College of London has revealed that people can only recognize artificially generated speech in 73% of cases. This result holds true for both English and Chinese languages.
The findings have prompted scientists to undertake the crucial task of developing automated deepfake detectors in order to combat fraudulent audio and video content.
Deepfakes are synthetic media that impersonate the voice or appearance of real individuals, and they fall under the category of generative artificial intelligence (AI). These AI systems are trained using media-based dummies.
Previously, the recreation of human speech required thousands of samples, but with modern algorithms, a 3-second voice recording can now be replicated.
The study employed a text-to-speech conversion algorithm trained on two public datasets in English and Chinese. The program generated 50 speech samples in each language that were different from the ones used in the algorithm’s training.
Of the 529 participants, only 73% were able to recognize the fake speech. Minimal improvement was observed even with short training on identifying deepfakes.
Kimberly May, the first author of the research, highlighted, “We used samples created by relatively old algorithms. The question arises – will people be able to notice anything suspicious in recordings generated by the latest technology in the future?”
While generative AI has positive applications for individuals with disabilities, there is concern among scientists that governments and criminals may exploit its capabilities.
An incident in 2019 involved scammers successfully emulating the voice of a British energy company’s CEO and convincing the company’s director to transfer large sums of money to them.
Recently, Apple announced the release of an application for iPhone and iPad that allows users to create a copy of a person’s voice based on a 15-minute recording.
Lewis Griffin, another author of the study, stated, “As the technology improves, organizations should develop strategies for control while still considering the advantages of these technologies.”