Neural Noise generates music snippets from the output of a Recurrent Neural Network (char-rnn). The networks (there are multiple to choose from below) were trained on a few thousand pop songs and take input and output in abc notation. You can read more about Neural Noise, or browse the open source code for this project on Github.
Since it takes upwards of 10 seconds to generate a song, and most visitors are interested in a random song anyways, the current function of this site is to return pre-generated songs at given temperature values.
As of 2015-08-20 there are approximately 500 songs per temperature value for checkpoints lm_lstm_epoch19.46_0.4127.t7 (non-real pop songs removed) and lm_lstm_epoch17.65_0.7648.7 (using nn format). The checkpoint lm_lstm_epoch21.83_0.3838.t7 (First attempt) has about 100 songs per temperature.
Neural Noise is a project of Travis Briggs. See more at https://travisbriggs.com
Select the name of a checkpoint and the temperature value to browse song snippets generated using that checkpoint at that temperature. Andrej Karpathy, the author of the char-rnn program that Neural Noise is based on, has a good discussion of temperature in his original blog post .