Do you believe it? Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai Ai.

In this AI created by MIT, researchers use a data set of millions of movie clips to self train a neural network model called speech2face. The operation of this network is roughly divided into two parts. One is a speech coder, which is mainly responsible for analyzing the input speech and predicting the relevant facial features The other is the face decoder, which integrates the input facial features and generates images. From the final results, only 6 seconds, we can rely on the voice to restore the face, the effect is satisfactory.

The purpose of speech face model is to study the relationship between speech and appearance?

The team said that their goal is not to accurately restore the appearance of the speaker, but to study the relationship between voice and appearance. At present, speech2face has been able to identify gender, and it is also easy to distinguish Caucasians and Asians. In the age range of 30, 40 and 70, the voice hit rate will be higher.

In addition to the basic gender, age and race, speech2face can also guess some facial features, such as nose structure, lip thickness and shape, occlusion and so on. It can also guess the general facial skeleton. Basically, the longer the voice input time is, the higher the accuracy of this AI will be. However, the researchers also admit that there will be mistakes in AI’s hearing, and this AI will not be experienced In the voice changing period, young boys are regarded as women, and they may misjudge the speaker’s accent, or even make mistakes in age. Researchers say that the limitation of spech2face is partly due to the fact that the speakers in the data set are not rich in ethnic diversity, so their ability to recognize the voices of different ethnic groups is relatively weak.

However, some people think that the privacy and discrimination hidden behind this technology are worrying; they think that although this is a pure academic investigation, the potential sensitivity of facial information is the moral factor in further discussion, and it is necessary to conduct strict technical tests to ensure that the actual data can represent the expected user group.

Editor in charge: PJ

Leave a Reply

Your email address will not be published. Required fields are marked *