Facebook Music Neural Network learned to fake singers voices

Scientists from Facebook Ai Research, together with developers from the University of Tel Aviv, created a musical neural network, which can convert audio record under the style of singing any artist.

It was possible to achieve the result thanks to the Google-based autocoder — a neural network, which uses training without a teacher. The autocoder generates a voices model of any artist based on the array of his songs. Then he compares the characteristics of the voice of any person with the timbre, height, force, tone and the tempo of the speech of the generated model and customizes them under the singers. This happens until all the characteristics of the compared audio records are the same.

The use of artificial intelligence in the musical sphere can help not only vocalists, but also people with speech defects. Evaluation of the system was carried out on the basis of audio recordings of 17 singers and 14 songs. Independent reviewers rated the similarity of transformed records with the voices of singers for 4 out of 5.

Artificial intelligence is gradually being introduced into all areas of creativity. Recently