Fact check: How do I spot audio deepfakes?

Did former US President Barack Obama suggest that the Democrats were behind the failed assassination attempt of his successor, Donald Trump?

Several audio recordings have been circulating in the US, ostensibly of Obama speaking with his former advisor, David Axelrod, about the upcoming US presidential election in November.

In one of the snippets, a voice that sounds like Obama's says: "It was their only opportunity and these idiots missed it. If only they could get rid of Trump, we would ensure their victory against any Republican candidate."

But the audio has been identified as fake, meaning Obama never said any of this. Instead, the audio was synthetically generated with the help of artificial intelligence (AI).

NewsGuard, a misinformation and media watchdog in the United States, published an analysis of the audio files. It used multiple AI detection tools and interviewed a digital forensics expert before concluding they were fake. It also spoke with an Obama spokesperson, who confirmed they were not authentic.

Obama is not the only politician with deepfakes of their voice out there. Earlier this year, an AI-generated audio of US President Joe Biden's voice urged voters in the New Hampshire primary election not to vote.

And it's not just a problem in the United States. Last year, shortly before an election in Slovakia, an audio deepfake went public impersonating liberal party leader Michal Simecka. In the UK, London Mayor Sadiq Khan also fell victim to a fake AI recording of him supposedly making controversial remarks.

Audio deepfakes have become a significant disinformation threat, especially in times of political uncertainty, such as during elections.

Easier to make, harder to debunk

AI-generated audio fakes can be particularly harmful in election cycles because they are so easy to create and disseminate.

"They require less training data and computing power — compared to video deepfakes — to produce nearly realistic outcomes," said Anna Schild, an expert in media and communication in DW's Research and Cooperation Projects team.

She has been examining the impact of audio deepfakes , together with her colleague Julia Bayer, and explains why they are fast-growing in popularity. 

"Their versatile applicability, from robocalls to voice messages and video voice-overs offer many different dissemination channels," Schild said.

Audio deepfakes are also harder to detect than other forms of disinformation.

"Audio fakes are a bit more difficult to recognize than video deepfakes because we simply have fewer clues," Nicolas Müller, a machine-learning engineer at Germany's Fraunhofer Institute for Applied and Integrated Security, told DW.

"In a video, we have the audio, the video, and a certain synchronicity between them," said Müller, who has studied people's abilities to detect fake recordings. He and his colleagues found that, in audio files, there are fewer elements people can rely on to detect whether the recording is authentic or not.

So, what can users do if they encounter an audio file that they feel may have been created by AI?

One solution would be to combine standard verification techniques with the use of AI software specialized in identifying audio deepfakes.

Honing the senses

One way to verify whether audio recordings are real or not is to check the file for telling patterns that could indicate AI interference.

In the above-mentioned example of Barack Obama, that would mean comparing the suspicious file with a known and verified audio track of his voice to find any possible deviations from Obama's normal manner of speaking.

This could include differing pronunciations, unnatural pauses, or unrealistic breathing patterns.

Beyond this, a closer look at the audio in question could involve checking for background noise or unnatural sounds.

Finding these clues can be hard for untrained listeners, but several tools have been designed to help people practice recognizing this kind of disinformation.

One of them is the Digger deepfake detection project, designed in cooperation with DW. The project has developed practical exercises for people to train their critical listening skills.

Nicolas Müller's team has also developed a game for participants to test how well they can spot audio deepfakes.

Using AI tools to combat AI disinformation

An additional verification layer involves using AI-supported software trained to detect audio deepfakes.

In our example with the synthetic voice of Obama, NewsGuard used deepfake checkers like TrueMedia , which has a deepfake detector bot that it says can reply to users' verification requests on the social media platform X (formerly Twitter).

Fraunhofer, meanwhile, developed Deepfake Total, a platform where users can upload suspicious audio files to have them analyzed. All uploaded files are rated with a score on a "fake-o-meter," which indicates the likelihood of the file being artificial.

It’s important, however, to stress that tools to detect deepfakes are not infallible. While they can estimate the likelihood of a file being AI-generated, they are not always correct. That is why such tools should always be used with caution, as one of multiple verification steps.

This could include browsing fact-checking sites and platforms to see if the audio in question has already been debunked by other fact-checkers.

Several media outlets have also developed tools to identify audio deepfakes, such as VerificAudio, by Spain's global media company PRISA Media, which aims to detect fakes in the Spanish-speaking world.

Jose Gutierrez from PRISA Media explained to DW that the tool is based on two AI-guided processes: While the first compares suspicious audio files with authentic audio recordings of the same person, the second analyzes acoustic features such as bandwidth, pitch, frequency and sound texture.

Gutierrez also stressed that this tool did not provide definitive answers, but percentages of plausibility.

Checking the context

If all this seems too complicated or technical, what also helps while trying to identify audio deepfakes is to focus on more traditional verification skills that are not exclusive to audio recordings.

DW's Research and Cooperation Projects team suggests "zooming out" to check content, source and any other relevant information on the file in question. They list some working tools on a website called "How to Verify".

Some helpful tips include comparing the audio content with known facts, checking the person’s social media channels, and searching for additional context on trustworthy news sources. 

In the end, it's all about using a mixture of techniques. As DW Research and Cooperation Projects points out, "there's no one-button solution that can assist and detect any kind of manipulation in audio."

This article is part of a DW Fact Check series on digital literacy. Other articles include:

• How do I spot manipulated images?
• How do I spot AI-generated images?
• How do I spot a deepfake?
• How do I spot state-sponsored propaganda?
• How do I spot fake social media accounts, bots and trolls?

And here you can read more about  how DW fact-checks fake news.

Edited by: Rachel Baig