Blog to post each day a RNN generated music
HTML CSS Python JavaScript
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
.sass-cache
_includes
_layouts
_posts
_sass
css
img
js
.gitignore
404.md
LICENSE
README.md
_config.yml
about.md
create_post.py
favicon.ico
feed.xml
index.html

README.md

RNN Music of the day

This is a blog in which we post everyday a music composed by a trained Recurrent Neural Network (RNN).

This is just for fun. I am not claiming this music to be superior or even equal in quality to human produced music. I'm just amazed about how good the results are given how simple the method is, so I wondered that some of you might find this cool.

The idea started some months ago, when Andrej Karpathy posted in his blog the article The Unreasonable Effectiveness of Recurrent Neural Networks, in which he talks about training a RNN to generate text a character at a time after being trained on some dataset. In that article he shows some examples, including Shakespeare like generated texts. He also released it's code.

Soon enough the Internet was flourishing with amazing things created using his code. Some examples:

The last two entries above are very similar to what I did, but I still think there is some merit in my work, such as:

  • I trained the network in much more data (8 MB versus 16 KB vs 415 KB) with greater musical variety.
  • I synthesized the results in better sounding mp3 files.
  • I'll be posting results daily here in this blog.

Let me give you some more details.

To train the network I started to look for ASCII notations for music and I found the ABC notation. It's a format much more suitable for training the network than common MIDI representations, where each track is too far apart for the network to correlate them. Another nice thing is that it supports chords and the chords are also contiguous to the melody in the same bar. On the other hand, it's used mainly by the folk music community, so the network was trained mainly with this kind of music, but from many different sources and types, including traditional Irish, Swiss, English, French and Chinese tunes. I collected a relatively huge amount of music in this notation to allow the network to get more variety and better generalization.

I processed the results keeping only the valid files, valid here meaning without errors or warnings being thrown by abc2midi. In this way I was able to generate hundreds of musics which I will release here.

For the more curious, I've released the scripts I used with some guidelines on how to use them. Feel free to ask questions.

Finally, I hope you enjoy this experiment's results as much as I do!

License

The MIT License (MIT)