Scarily, AI Can Generate Accurate Faces Just From a Person's Voice

Scientists at the Massachusetts Institute of Technology (MIT-USA) have for the first time successfully applied an algorithm to recreate a person's portrait using only a short voice recording.

The AI ​​algorithm called Speech2Face was first introduced in 2019.

First, the researchers designed a deep learning artificial neural network. The AI ​​was then trained by watching millions of videos from YouTube and the internet of people talking to learn the correlation between the sound of a voice and the speaker, and then make its best guesses about the speaker's age, gender, and nationality.

Once trained, the AI ​​was able to come up with portraits based on voice recordings alone.

The researchers built a 'face decoder' that creates a standard representation of a person's face from a still image of them, ignoring lighting and pose. They then used this standard human face to compare it to the AI's voice-generated face. The results showed that the AI-generated face was very close to the real face in a wide range of studied cases from a variety of ages, genders, and ethnicities.

AI-generated portraits could be used to assign machine-generated voices to home appliances and virtual assistants, researchers say. Or AI could help law enforcement create a portrait of a suspect from a voice recording as the only evidence. But that could raise privacy concerns.