Voice deepfakes: how they are generated, used, misused and differentiated

Last Updated08/02/2023

Recently several users of the social media platform, used “speech synthesis” and “voice cloning” service provider, to make voice deepfakes of celebrities.
These deepfake audios made racist, abusive, and violent comments.
Making deepfake voices to impersonate others without their consent is a serious concern that could have devastating consequences.

A voice deepfake is one that closely mimics a real person’s voice.
The voice can accurately replicate tonality, accents, cadence, and other unique characteristics of the target person.
People use AI and robust computing power to generate such voice clones or synthetic voices.
Sometimes it can take weeks to produce such voices, according to Speechify, a text-to-speech conversion app.

To create deepfakes one needs high-end computers with powerful graphics cards, leveraging cloud computing power.
Powerful computing hardware can accelerate the process of rendering, which can take hours, days, and even weeks, depending on the process.
Besides specialised tools and software, generating deepfakes need training data to be fed to AI models.
This data are often original recordings of the target person’s voice. AI can use this data to render an authentic-sounding voice, which can then be used to say anything.

Attackers are using such technology to defraud users, steal their identity, and to engage in various other illegal activities like phone scams and posting fake videos on social media platforms.
Voice deepfakes used in filmmaking have also raised ethical concerns about the use of the technology.
Gathering clear recordings of people’s voices is getting easier and can be obtained through recorders, online interviews, and press conferences.
Voice capture technology is also improving, making the data fed to AI models more accurate and leading to more believable deepfake voices.
This could lead to scarier situations, Speechify highlighted in their blog.

Open AI’s Vall-e, My Own Voice, Resemble, Descript, ReSpeecher, and iSpeech are some of the tools that can be used in voice cloning.
ReSpeecher is the software used by Lucasfilm to create Luke Skywalker’s voice in the Mandalorian.

Detecting voice deepfakes need highly advanced technologies, software, and hardware to break down speech patterns, background noise, and other elements.
Cybersecurity tools have yet to create foolproof ways to detect audio deepfakes.
Researchers developed a technique to measure acoustic and fluid dynamic differences between original voice samples of humans and those generated synthetically by computers.
They estimated the arrangement of the human vocal tract during speech generation and showed that deepfakes often model impossible or highly unlikely anatomical arrangements.