The Magenta team is very proud to have been awarded “Best Demo” at the Neural Information Processing Systems conference in Barcelona last week.
Here is a short video of the demo in action at the Google Brain office:
The demo consists of several components:
- The Magenta library, which is built with TensorFlow and provides an API to to generate MIDI data with neural network models.
- The Magenta-MIDI interface, which provides an interactive communication layer between TensorFlow and MIDI devices.
- Models developed by the Magenta team and our collaborators, including a set of 6 neural networks: a basic one-hot melody RNN, a melody RNN with lookback, a melody RNN with attention, an RNN which had been fine-tuned with reinforcement learning, a drum RNN and a polyphonic RNN based on BachBot. Pre-trained versions of these models can all be downloaded on our GitHub.
- A visualization built by Google CreativeLabs.
- A MIDI keyboard and drum pad.
- An Ableton Live / MaxMSP interface that connects all of the above, so that participants can improvise with the models.
We’re also pleased that we can provide you with all of the code we wrote to put this demo together. Instructions and code for running the full demo, which requires Ableton Live and MaxMSP, are now available in our GitHub. We have also created a JavaScript version of the demo that isn’t dependent on these expensive software packages.
If you use this interface to produce music, aid a live performance, or even in your own AI jam session, we’d love to hear about it! Please drop us a line at magenta-discuss@google.com.