DeepComposer By Amazon: First Neural Network Music Synthesizer

Amazon introduced the DeepComposer music synthesizer and the eponymous cloud-based music creation service based on generative adversarial neural networks. Using them, the user can set the main melody on the synthesizer and get a full song, in which the original part is supplemented with drums, guitar and other instruments.

Researchers in the field of machine learning have long been working on creating algorithms that can help people create works of art or even do this work on their own. Machine learning algorithms working with images, for example, changing the style of paintings, received the greatest attention.

A lot of researchers also work in the field of neural network synthesis of music, and some of them, for example, the Magenta group as part of Google AI, have achieved considerable success, including in applied developments that musicians can potentially use. For example, they created an algorithm that complements the game on any instrument with a corresponding batch of drums.

Amazon introduced something similar, but wider and more ready for mass
use - DeepComposer service. This service operates on the basis of neural
network algorithms that obtain a sequence of playing a single instrument and synthesize, in addition to it, sequences for other instruments.

So far, DeepComposer can complement the original sequence with several instruments, such as an electric guitar, bass, drums and synthesizer. In addition, the user selects from trained models for rock, pop, jazz and classical music, however, they can upload their own data set and train the model for the desired style.

Together with the DeepComposer service, the company introduced a
compatible 32-key synthesizer. In addition to the main keys, it is equipped with 11 function keys, for example, an octave change button. It connects to a computer via USB and transmits data in the form of a MIDI sequence.

The cost of the keyboard is $ 99, and the cost of the service is calculated from the operating time of the servers when training and using neural network models.