Week 1 (2/28) – Picking Out Pieces
Edward W -
This week’s goal: Pick out all of the pieces that I plan to use in the resulting survey.
A few setbacks first…
So… it turns out that the MuseNet neural network I was planning to use for my senior project is on hiatus indefinitely, and there is no official MuseNet neural network that has been released by OpenAI. The research for the neural network has been on pause since December 2022, but I didn’t realize this, so I was searching on GitHub for nearly an hour until I found a post saying the neural network was no longer accessible to the public. Although it does mean that I have to pivot to using another neural network, it doesn’t affect any of my progress this week, since my objective this week was to pick out pieces to use in my survey.
But currently, I am looking at Allegro Music Transformer, which is a neural network that functions similarly to MuseNet, but has all of its code accessible on asigalov61’s GitHub page. Allegro Music Transformer also uses Jupyter Notebook, which I am very new to, so it’s a bit of foreshadowing to how next week is going to go…
Back to this week…
This week was spent combing through piece after piece by composer after composer. I think I’ve listened to almost a good eight hours’ worth of music in the past week, and this might constitute the most boring part of the project. The good news is that I have a set of ten pieces picked out, and although I can’t specify what pieces I will be using in the final survey, I can tell you what genres and arrangements I’ve picked!
- Baroque – solo organ, four-part fugue
- Classical – string quartet
- Romantic – solo piano
- Contemporary – solo piano
- Contemporary – solo cello
- Jazz – jazz quartet, sax solo
- World – vocals over lyre
- City Pop – pop band (vocals over synth, bass guitar, and drumset)
- Ambient – piano and synth
- Original Composition – piano and cello duet
I picked a lot of piano and cello pieces, mostly because that is what me and my piano teacher are familiar with, but I also tried to include other ensembles that I am less familiar with, like jazz quartets, pop bands, and organ.
Although I would have liked to include more music from the modern era, the problem with pop music is that… well, it’s popular! For any pop song I pick, there is a decent chance that one of the respondents might actually know the song, and it ends up ruining the point of the survey. Now that I have these pieces, I am going to have to transcribe them into MIDI files that can be processed through MuseScore, which will be coming up on Week 4.
Below is a picture of me listening to music! (Here, it’s Mozart’s Ein musikalischer Spaß.)
Next week’s objective
Next week is the fun part! I’ll be figuring out how both Allegro Music Transformer and Jupyter Notebook work, and I’ll be starting to actually create music using neural networks. For now, the pieces that I found will give me a good baseline for what genres I should try to get the AI to compose in.
Thanks for sticking around, and hopefully I’ll see you in next week’s blog post!
– Eddie 😛
Minor Edit (3-4-24)
I wanted to update this post, especially since I think Isabelle asked a very good question about what qualities I was looking for while I was choosing my pieces, and I wanted to take the opportunity to address it here since I should’ve but didn’t in this post last week.
I was looking for three major factors in all the pieces I was picking:
- Recognizability – A piece is recognizable when I can give it to someone with a lower level of musical experience, and they can go, “Oh, it’s this genre” or “Oh, it’s from this musical era”, so it was very important that these pieces are representative of their genre or era.
- Obscurity – A piece is obscure when it isn’t as well known in the musical community. Of course, I can never guarantee that absolutely no one has heard of a specific piece, but because of the sheer quantity of music many composers put out, there are many pieces that just never gained popularity. The last thing I would want to happen is me using a common piece like Für Elise or the Flight of the Bumblebee and have respondents to my survey be able to deduce which piece they mark as AI-generated using this sort of deduction.
- Arrangement Commonality – I needed to pick pieces that use common ensembles. For example, if I used one of Liszt’s piano duets (of which he has only three), a listener could potentially tell that it’s Liszt not from the content of the music, but by which instruments are used, which defeats the point of the survey.
Here are some examples of the first factor, since it might not be completely clear how music from the Romantic era may not sound “Romantic” in style, etc.:
- Ludwig van Beethoven’s Sonata in C-minor, No. 32, Mvt. II – Arietta – youtu.be/WGg9cE-ceso?t=955 – This section feels like it could fit in snugly with 1920s boogie-woogie, not the 1820s from whence it came!
- Heinrich Biber’s Battalia à 10, Mvt. II – Die liederliche Gesellschaft von allerley Humor – youtu.be/5YBOmgi-qSs?t=105 – The movement is very reminiscent of contemporary music, despite being from 1673, containing dissonances that would make Schönberg blush!
- Franz Liszt’s Bagatelle sans tonalité, S. 216a – youtu.be/yc_HjEa8k5k – The whole piece is filled with atonality that is reminiscent of contemporary classical music, and in a way, Liszt cleverly predicted the direction that classical music would go in in the years following his death.
This time it has working links too!
I wanted to add this because I regret not adding into my original post, but I think it’s definitely important context for readers to know!
Thanks again to Isabelle for the question!
Comments:
All viewpoints are welcome but profane, threatening, disrespectful, or harassing comments will not be tolerated and are subject to moderation up to, and including, full deletion.