Initial Conclusion of Robots Learning to Play Drums
or Rhubarb's Triumph
15 May 2021 - atrodo - Song: Mother's Triumph by Matthew Thiessen & The Earthquakes
It has been 30 days since I began releasing music created by robots. It’s be a very interesting 30 days, and I’ve learned quite a lot doing it. I’ve already made and planned a couple improvements for the existing set of uploads and am also looking at future improvements. With all of that that behind me, I wanted to record briefly what I did do and some ideas and plans I have for improvements going forward.
Initially I was hoping to do a larger, more comprehensive campaign around posting to social media while analysing things that were happening and making course corrections as time went on. I ended up deciding to do no social media posts nor any major structural changes to the content, like changing video items, thumbnails, or the like. Instead, I focused on establishing a baseline of what to expect in the future with a few mid-course changes to the music generation process.
The biggest mid course change was an adjustment to the data I fed to the neural network. Early on I was using a basic on/off value to denote that the melody is playing a note, with an additional value to indicate that the note is about to finish to help the robots find the end of notes. The biggest mid-stream change I decided to do was adding more data and use a sawtooth pattern between 12 values, high to low. This was to indicate which subbeat of the melody note is currently playing. This gave the drummers a much better opportunity to have short staccato hits.
Then I discovered that my normalization for notes was not giving the robots an accurate view of notes; notably they were being given weird values when the note wasn’t near Middle-C. I adjusted that so it is given a more honest view of the note being played, and corrected the output so it was actually related to the circle of fifths and not just made up notes with no relationship.
Going forward, I have a list of ideas of change after listening to a couple things. That doesn’t give me a complete picture of the current progress, so I’m going to earnestly listen to all 30 videos, take notes, and rate how much each change could possibly improve it. I am hoping that in the process I’ll come up with more ideas and a solid go forward plan.
Overall, I’m happy I did this 30 day run. It got me real world experience with what is needed to process this music on an on-going basis and what I can expect if I’m doing daily videos. I’m planning on working methodically with these changes, as well as taking a small break to handle some bits that have fallen to the wayside. I hope that in a few months I’ll be able to restart daily videos.