Humans Unable to Accurately Identify Deepfake Voice Samples

2023-08-03

A new study from University College London has found that humans can only identify artificially generated speech in 73% of cases, with equal accuracy for English and Mandarin Chinese. This study, published in PLOS ONE, is the first to assess human ability to detect artificially generated speech in languages other than English. Deepfakes are a form of synthetic media that can mimic the voice or appearance of a real person. They fall under the category of generative artificial intelligence, which is a type of machine learning that uses trained algorithms to learn patterns and features from datasets, such as videos or audio recordings of real people, in order to recreate their voices or images. While early Deepfake algorithms may have required thousands of voice samples from different individuals to generate original audio, the latest pre-training algorithms only need a three-second clip of a person speaking to reconstruct their voice. Open-source algorithms are freely available, and with some expertise, individuals can complete the training process in a matter of days. Recently, Apple released software for iPhone and iPad users that allows them to create their own voice clones through a 15-minute recording. Researchers at University College London used a text-to-speech (TTS) algorithm trained on two publicly available datasets, one in English and one in Mandarin Chinese, to generate 50 Deepfake speech samples for each language. These samples were distinct from the ones used to train the algorithm, in order to avoid the possibility of the algorithm replicating the original input. 529 participants listened to these artificially generated samples and real samples to determine if they could distinguish the genuine from the fake. Participants were only able to identify 73% of the fake speech, and this percentage only slightly improved after receiving training specifically focused on identifying deepfake speech. Kimberly Mai, the first author of the study, stated, "Our research findings confirm that humans cannot reliably detect deepfake speech, regardless of whether they have received training to help them identify synthetic content. It is also worth noting that the samples used in this study were created using relatively older algorithms, which raises the question of whether humans would find it even more difficult to judge deepfake speech created using the most advanced technologies available now and in the future." The next step for the researchers is to develop better automatic speech detectors to enhance detection capabilities and counter the threat of artificially generated audio and images. While generative AI audio technology has benefits, such as providing convenience for individuals with language limitations or those who have lost their voices due to illness, there is growing concern about the potential misuse of this technology by criminals, leading to significant harm to individuals and society. Recorded cases of criminals using deepfake speech include an incident in 2019, where the CEO of a British energy company was convinced to transfer hundreds of thousands of pounds to a fraudulent supplier by a deepfake-generated recording of their boss's voice. Lewis Griffin, the lead researcher of the study, said, "As generative AI technology becomes increasingly sophisticated, with many tools being publicly available, it brings both benefits and risks." Source: https://techxplore.com/news/2023-08-humans-unable-quarter-deepfake-speech.html

RECENT AI NEWS

RECENT AI TOOLS