Amazon, in addition to selling products online, has a huge (huge) web services and artificial intelligence division called AWS (Amazon Web Services). In it, Amazon brings together everything related to tasks in the cloud, such as voice to text transcription, augmented reality, blockchain and, of course, machine learning. It is precisely the latter that takes advantage of DeepComposer, the latest release of the brand that is nothing more than a musical keyboard to create new songs from a melody using machine learning.
According to Amazon, it is the first of its kind. It is a keyboard similar to a piano with 32 keys designed for developers “to have access to generative AI, either with pre-trained models or theirs.” At the moment the device is in the testing phase, but it can be used in the cloud through the AWS console in certain regions of the United States. Its operation is really simple in practice, although what happens behind the scenes is really interesting.
From a Beethoven song to a rock theme
How does DeepComposer work? For the user, he only has to log in to the AWS console, record a short melody or import an already recorded one, select a generative model corresponding to his favorite genre (say, rock) or import his own model and play the composition on the console Automatically, the AI returns a similar composition, but modified and adapted to the model we previously chose that we can import directly into SoundCloud.
Behind, what we really have is an antagonistic generative network (GAN), which consists of a generating neural network and another that discriminates. The generator uses random data to create samples that are sent to the discriminator, and it must identify whether the sample is part of the training corpus or if it is a false sample created by the generating neural network. Gradually, the generating network learns to create compositions so real that the discriminating network is not able to identify as false.
What’s the score?
Below you will see two audios. The first is a 20-second fragment of Beethoven’s ‘Ode to Joy’. The second is the result of asking the antagonistic neural network to create a 20-second composition of the same theme using the pre-trained “Rock” model. The AI has not only adapted the song to the tempos of a rock song, but has been able to add drums, guitar, electric guitar and bass.
As we said at the beginning, DeepComposer can be purchased in physical format, although it can be used on the web from some regions of the United States. Its price will be 99 dollars and, at first, it stays in its country of origin. According to Amazon, “availability for other selected countries will be announced later.”
In the case of opting for the web version, Amazon offers 12 months of Free Tier (that is, basic use). With this plan we can create a piece using one of the pre-trained models and import a melody to create up to 500 songs. In addition, a free 30-day AWS DeepComposter trial is offered to all users with which up to 40 themes can be created. Subsequently, the price will become $ 1.26 per hour of training and $ 2.14 per hour of inference.