Member Login
menu

This 'DrumNet' AI Can Play Drums After Learning Patterns Of Rhythmic Interaction

When researchers know that AIs can be taught to be play music, they have created many "musicians".

Among others, there is an AI capable of playing Death Metal to one that plays Jazz. Not to mention one AI that composes a Country song.

Things don't stop there, as Sony Computer Science Laboratories (CSL) in Paris, France, also wanted to experiment with AI by developing one that plays drums.

Called 'DrumNet', the AI can add some kick-drum beats to pre-existing songs, and make them more interesting.

Below are four originals created by the AI Headphones or good speakers are required to the examples, as the kick drums may not be audible otherwise.

Drehscheibe:

Orgs Waltz:

Gipsy Love:

Miss You:

According to the researchers' paper:

Here we propose a model for conditional kick drum track generation that takes existing musical material as input, in addition to a lowdimensional code that encodes the desired relation between the existing material and the new material to be generated.

Using unsupervised learning of music dataset that consists of 665 Pop, Rock and Hip-Hop tracks, the AI is taught to learn patterns from rhythm instruments that include bass, kick and snare, which were all available as separate 44.1kHz audio tracks.

The researchers show that codes can indeed be sampled to create a variety of musically plausible kick-drum tracks, in which the model can be used to transfer kick-drum patterns from one song to another.

The result, is the AI in adding "musically plausible" beats invariant to tempo and time-shift.

DrumNet
The proposed model architecture

In other words, DrumNet that is based on AI neural network, leaned how to play this klck-drum by understanding the patterns of rhythmic relationships between different instruments.

Detailing the tracks as the sequence of mapping code of kick vs. bass and snare inputs, the researchers said that:

"Each mapping code captures local relations between kick vs bass and snare inputs, such that an entire track is associated to a sequence of mapping codes. Since we want mapping codes to capture rhythmic patterns rather than just the instantaneous presence or absence of onsets in the tracks, during training we enforce invariance of mapping codes to (moderate) time shifts and tempo changes in the inputs."

After all the processing, researchers made the AI to play drum-kicks by placing a ‘one-shot’ drum sample at amplitude peaks.

The AI adjusts the generated beats in relation to existing tracks in a 16-dimensional style space, to make it sound as it it plays drums.

"Although the current work is limited in the sense that the model has only been demonstrated for kick drum track generation, we believe this approach is applicable to other content," said the researchers.

But here, "the goal is not to replace musicians, but to provide them with better tools to be more efficient in realizing their creative ideas."

Published: 
07/08/2019