How much can you infer a person's appearance based on their speech? That is the question with which a group of researchers from the Massachusetts Institute of Technology (MIT, for its acronym in English) began a project that aims to make an algorithm capable of generating only the most characteristic physical features of a person. with his speech. As a result, they created Speech2Face, an AI that manages to create a virtual face very similar to that of a human only by reproducing a few seconds of audio with their voice.
In this way, the algorithm can have a multitude of references that allow it to create a face without the need for any image.
Some AI-generated faces are even more trustworthy than real ones
Real faces and created by AI (S).
Unlike the AI called Speech2Face, which can create an avatar of a person just by listening to the voice, there are machine learning systems capable of generating faces so similar to those of a real person that even humans cannot differentiate. An investigation published in the journal PNAS, in fact, shows that the majority of people who participated in a study whose objective was to check if they could distinguish real faces from virtual ones, found those generated by artificial intelligence more reliable than real ones.
These types of virtual faces are created using a system made up of two neural networks. The first is responsible for generating the avatars from a matrix of pixels. The second, on the other hand, is dedicated to comparing the face created by AI with a real one. If it finds differences, it penalizes the first system so that it does not make the same mistake again. Thus, until the image is extremely similar to the real one.
These learning systems can have significant benefits if they can be applied on a day-to-day basis. For example, AI capable of creating a face from voice can be useful for generating avatars of criminals. However, they also pose a significant danger. Above all, considering the ease with which a face can be created to, among other uses, impersonate a person's identity.
