Foto: Unsplash

“You trust a voice”

Tuesday 01 Dec 20
by Christina Tækker, Jesper Spangsmark


Morten Mørup
DTU Compute
+45 45 25 39 00

Students from Artificial Intelligence and Data would like ethical guidelines after testing advanced models for voice conversion.

When a group of second-year students in the BSc programme Artificial Intelligence and Data defended their course project ‘Deep Voice Conversion’ at the exam, they surprised both the external examiner and their supervisor when they falsified former US President Barack Obama’s voice so convincingly that it sounded like President Donald Trump.

Voice conversion—or deep voice conversion—is a type of deepfake in sound and is produced using artificial intelligence. Deepfake is an overall concept that includes sound, text, and images. Using neural networks that can learn complicated contexts from large data volumes, you can distort a person’s face so that his or her facial expressions and gestures match another soundtrack—just as you can fake another person’s voice.

“There has been a lot of media coverage of deepfake and fake news, and how you primarily manipulate images and create fake videos. But voices can also be manipulated and stolen. We’re seeing more and more examples of this in politics and entertainment, among other fields,” explains Peter Grønning, who has prepared the course project together with Gustav Gamst Larsen and Lukas Leindals.

Dilemmas and grey areas

The purpose of their project was to identify the latest models for voice conversion, compare them, and consider the ethical dilemmas and grey areas that may arise using this technology. But the main part of the project involved testing the models and experimenting with their own voices and those of friends.

“We examined how far we could come with two years of code experience, and we discovered that it’s difficult to make a good, convincing deep voice conversion. But voice conversion seems to be becoming increasingly accessible and has great potential—also to con a lot of people,” says Lukas Leindals.

"Voice conversion seems to be becoming increasingly accessible and has great potential—also to con a lot of people."
Lukas Leindals, Student on Artificial Intelligence and Data

Together with his fellow students, he would like to see ethical guidelines on how to use the technology. This applies both to guidelines for how to keep data anonymous and how to use data, such as informing the owner of the voice about what the voice will be used for.

“Your voice is an essential part of your identity. You trust a voice. If I hear my mother’s voice, I have no doubt it’s her,” says Peter Grønning.

However, the technology can be used for other things than manipulation. It also offers good applications, as it can be utilized to give people a voice, says Gustav Gamst Larsen:

“Deep voice conversion can help people who have lost their voice. Instead of a mechanical robot voice, you can reconstruct your own voice—and thus improve your quality of life.”

Many young people want to study AI

DTU’s BSc programme in Artificial Intelligence and Data is among the study programmes in highest demand at the University. 

There is great interest in studying artificial intelligence (AI) at DTU. In 2018, DTU set up the BSc programme Artificial Intelligence and Data, which has become one of the study programmes in highest demand at DTU.

The initiator of the new study programme was Professor Lars Kai Hansen, who stated in connection with the launch:

“Artificial intelligence will be a central part of our future—and graduate engineers from the study programme ‘Artificial Intelligence and Data’ will be the digital designers of this future. They will gain an insight into the principles behind learning algorithms, and how computers make decisions and communicate with users. They will also be able to equip computers with the skills that lie at the core of our own intelligence.”

120 places offered in 2020

As the development of artificial intelligence is closely linked to human behaviour and human values, the students will also acquire knowledge of cognitive psychology and social mechanisms as a basis for designing and evaluating intelligent systems—also in relation to ethics and privacy.

From offering 45 places in 2018, DTU offered 120 places on Artificial Intelligence and Data in 2020.

AI engineers are especially in demand in the pharmaceutical industry, manufacturing, banking, IT, and the transport sector, and many new start-ups are employing engineers with AI competences.