innovator

Add idea


Calendar

«    May 2018    »
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: 24-05-2018, 06:42
Researchers from Facebook Research have developed a method for converting music records into recordings with other instruments, style and genre. This method is based on machine learning without a teacher and allows you to train a neural network without comparing records with different styles, instruments or genres. The developers demonstrated the possibilities of the method on different musical styles and artists, for example, turning Mozart's symphony into a recording of a pianist playing Beethoven's music. Work with the method description is published on arXiv.org.
The authors of the work have trained the neural network on the music of Mozart, Haydn, Bach and Beethoven. After training, the researchers conducted several experiments to test the operation of the neural network. In one of them, they invited three unnamed experienced pianists, who were to convert 60 five-second segments of musical records into piano music. One third of them were records of Bach's works for keyboard instruments, one third by Mozart's symphonies, and another third were pieces of swing jazz songs, guitar riffs and Chinese instrumental music.
The same thing was done by the neural network.
As a result, the neural network received lower estimates for the quality of the translation, but still managed quite well, especially when turning into piano music records from the latter group.
Note that the authors have shown many examples of the operation of a neural network. Including, in the video you can hear the transformation of the songs of the group Metallica, Elvis Presley and Rihanna into classical music.