Audio has only been tested to work in Chrome and Mozilla Firefox.

Additional Audio Samples

Below are samples from many different long short term memory (LSTM) recurrent neural network (RNN) models. Please visit the github repo for this project for more information about how these samples were generated.

Longer samples

Here’s a youtube video of a longer 30 second sample from the Corrette network.


The video below demonstrates the problem of overfitting that I initially encountered.



Audio

Warning: Some samples may sound horrible! Keep your hand close to the volume.

Sample of an untrained network

In this network, all of the feature weights are gaussian normal.

A sample of what the network “hears” during training

This is parsed snippet from an organ composition by Thomas Arne.

If you liked this you may also like this:



Network trained on Beethoven for ~100,000 epochs



Samples of network trained on organ music of Corrette



Sample of network trained on Romantic period music



Sample of network trained on Baroque period music



Sample of network trained on Classical period music



Images

Below are plots of predicted note probabilities as determined by the network on the x-axis and time on the y-axis. The band on the right-hand side is a one-hot encoding of the number of notes to play at that time, so density to the right indicates a louder part while density to the left indicates silence.

Mozart network

png png

Corrette organ music network

png

Beethoven network

png