Science & Technology, UK (Commonwealth Union) – A recent study conducted by the University College London has noted that humans were capable of just identifying just artificially generated speech 73 percent of the time, having the same preciseness in both English and Mandarin.
The latest research, published in PLOS ONE today, marks the first attempt to assess human capacity in detecting artificially generated speech in languages other than English.
Deepfakes are synthetic media designed to mimic the voice or appearance of real individuals. They fall into the category of generative artificial intelligence (AI), a type of machine learning (ML) that trains algorithms to learn patterns and features from datasets, such as audio or video recordings of actual people, allowing them to replicate the original sound or imagery.
Previously, early deepfake speech algorithms necessitated thousands of voice samples to generate authentic audio. However, the most recent pre-trained algorithms can recreate a person’s voice using just a brief three-second clip of their speech. Open-source algorithms are readily accessible, and while some expertise is advantageous, an individual could feasibly train them within a few days.
Tech giant Apple recently introduced software for iPhone and iPad, enabling users to create a replica of their voice with just 15 minutes of recordings.
In this study, researchers at UCL utilized a text-to-speech (TTS) algorithm trained on two publicly available datasets—one in English and one in Mandarin—to generate 50 deepfake speech samples in each language. To avoid the possibility of the algorithm reproducing the original input, these samples were distinct from those used for training.
The artificially generated samples, along with authentic ones, were played for 529 participants to determine their ability to distinguish between real and fake speech. The results showed that participants could identify fake speech only 73% of the time, and even with some training to recognize aspects of deepfake speech, their accuracy only marginally improved.
“Our findings confirm that humans are unable to reliably detect deepfake speech, whether or not they have received training to help them spot artificial content. It’s also worth noting that the samples that we used in this study were created with algorithms that are relatively old, which raises the question whether humans would be less able to detect deepfake speech created using the most sophisticated technology available now and in the future,” explained Kimberly Mai of UCL Computer Science who is also 1st author of the study.
The researchers’ next objective is to enhance automated speech detectors as part of their ongoing efforts to develop detection capabilities that can counter the threat posed by artificially generated audio and imagery.
While generative AI audio technology offers advantages, such as increased accessibility for individuals with speech limitations or those who have lost their voices due to illness, concerns are mounting over its potential misuse by criminals and nation states to cause significant harm to individuals and societies.
There have been documented instances of criminals employing deepfake speech, including a case in 2019 where the CEO of a British energy company fell victim to a deepfake recording of his boss’s voice, leading to the transfer of hundreds of thousands of pounds to a false supplier.
Professor Lewis Griffin from UCL Computer Science, the senior author of the study, emphasized that as generative artificial intelligence technology becomes more sophisticated and openly available, we are approaching a crucial juncture where numerous benefits and risks coexist. It is essential for governments and organizations to develop strategies to address the potential abuse of these tools. However, it is also crucial to recognize the positive opportunities that lie ahead.
Concerns over deepfake from across the world also centers around more research further enhancing the authenticity of deepfake as time goes by.
