innovator

Add idea


Calendar

«    May 2019    »
MonTueWedThuFriSatSun
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
 

 

Advert

 

Payment

 

Advert

 

Authorization

Стартап

DateDate: 29-05-2019, 05:58

Experts have taught technology to draw with the help of several million videos.
A group of American developers introduced the Speech2Face neural network, capable of drawing a portrait of a man according to his voice. Reported arXiv.org.
Models are able to draw an image, based on gender, race, age. The team engaged in the development of the Massachusetts Institute of Technology, which included Tahyon O.
Scientists decided to use AVSpeech for learning neural network. It contains about a million short videos of over a hundred thousand different people. In this case, the video and audio tracks are separated.
The accuracy of the technology makers determined according to three demographic indicators. Comparison was made of gender, approximate age and race of people from the original clips and the "conclusion" made by the neural network.
Were also identified the shortcomings of the model. She is not always able to determine age with an accuracy of ten years, and best of all depicts Caucasians and Asians. It is believed that the latter problem is related to the uneven presence of different races in the sample for training.
According to the researchers, their plans did not include a thorough copying of a person’s appearance. They sought to accurately identify gender, age, ethnic group.
In the photo: the original image, restored and "drawn" by voice.