Week 3 (3/20) – Composing My Own Music

Edward W -

This week’s goal: Compose all of my own music that I plan to use in the survey.

Hey, everybody! If you weren’t already aware, I took off last week for spring break, so in one of the upcoming weeks, I’ll try to squeeze in two posts instead of one, but I’ll still treat each one as its own separate week.

My own compositions…

I spent this week composing music, and I have to say… it’s definitely not my strong suit. I think this mostly stemmed from the fact that I had to take a seed and extend it, similarly to the AI. Usually, I am pretty comfortable composing my own original pieces, but it was definitely harder for me to replicate a style convincingly, especially when my compositional skill is nowhere near that of Beethoven or Mozart. So I wanted to preface what you’re about to hear with… I tried my best!

Needless to say, I can’t show you the actual pieces that I will be using in my actual survey, but I prepared some example pieces for you to enjoy! Because I have all the audio files for these pieces (the original ending to the seed, the AI-generated ending, and my own ending), you can have a sort of mini survey before the actual survey comes out (which I will be working on in Week 5).

For each piece, I give you the original seed that was given to the AI to generate and to me to compose, and an extra audio file that demonstrates how the seeds were continued by the composers who wrote them. I have randomized the original continuation, my composition, and the AI composition’s order like how I am planning to with the survey. Feel free to try and have a guess at it, and once you’re done, you can highlight the white space in front of the audio files to see how you did!

Also, feel free to let me know how the mini survey goes for you with the comments! I could always use more data to inform my hypotheses! Along with each composer, I’ll also go over some bumps that I ran into while composing and give a bit more insight into the pieces.

Without further ado…

A mini survey!

Bach, Johann Sebastian – Fugue 1 in C major from WTC I, BWV 846

Original Seed
> AI Batch #2
> AI Batch #1
> Original Ending
> My Continuation

Out of all of the compositions, this one was by far the hardest for me. Bach was one of, if not THE best fugal writer of his time, so trying to write a continuation to this piece was like learning how to drive from the backseat of a car. This was made worse by the fact that fugal writing consists of interlocking voices; each voice has its own independent melody, but they come together to form a complete composition. Each melody I wrote would sound fine on its own, but when it came together with the rest, it was a cacophonous mess, so I ended up spending hours adjusting each melody so that it could sound good with the whole.

Mozart, Wolfgang Amadeus – Eine kleine Nachtmusik, K. 525

Original Seed
> AI Batch #1
> My Continuation
> Original Ending
> AI Batch #2

This one was also a pretty big challenge for me. I am pretty unfamiliar with string quartet composition, so making the different instruments blend well with each other was a pretty large hill to climb over. Since Eine kleine Nachtmusik is so popular, I think you’ll probably be able to tell which one was Mozart’s continuation, but try and see if you can tell which two are the AI’s and which one’s mine!

Chopin, Fryderyk – Waltz No. 19 in A minor, Op. posth. B. 150

Original Seed
> AI Batch #2
> Original Ending
> AI Batch #1
> My Continuation

This one was by far the easiest to compose for. I am fairly familiar with Chopin’s work, especially his waltzes, so I think I was going into this one with more of an idea of what I could do to emulate the Romantic style that Chopin composed in.

Prokofiev, Sergei – Suggestion diabolique from 4 Pieces for Piano, Op. 4, No. 4

Original Seed
> AI Batch #2
> My Continuation
> AI Batch #1
> Original Ending

Being a Contemporary composer, Prokofiev has a pretty distinct musical style that I think both the AI and I had trouble replicating. I was very fixated on the chromatic motif that you hear at the beginning, as you’ll hear that throughout my composition. One other major problem I had was that I was using a lot of diminished chords and minor ninths, both of which are dissonant, to replicate Prokofiev’s sound, but Prokofiev seems to weave in consonance and dissonance throughout his pieces, which was really difficult for me to replicate, especially since I rarely compose in a Contemporary style.

W., Edward – Composition for tenor sax, bass guitar, piano, and drumset

Original Seed
> My Continuation
> AI Batch #1
> Original Ending
> AI Batch #2

This one was also an easier piece for me to compose off of, since I composed it myself! I think the main issue I had with this piece was that I have only a little experience with saxophone and jazzy writing, so writing a sax solo over the main chords was actually kind of difficult, but I really like the end result!

In conclusion…

I thought that this week would be more of an easy week, but I ended up spending a lot more time on these compositions than I am willing to admit. Composition is no joke, and this week gave me a lot more insight into how layered and detailed musical composition can be, even when it might sound pretty simple.

As always, below is a photo of me composing! (Or at least, trying to…)

Me composing music

Next week’s objective

Next week, I will be spending time transcribing all of my results through MuseScore, which is the sound font that I am planning to use for my final survey. This way, respondents to my survey won’t be able to distinguish AI and original compositions through different instrument sounds, because a lot of the time, different instruments sound different and render different in an audio file.

Thanks again for all the support, and I hope you stick around for the next one!

– Eddie 😛

More Posts

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.

    dain_y
    Hi Eddie! Wow, this blog post is so impressive along with the work you are doing! The survey was super fun, but I don't think I was qualified enough to make very good educational guesses, haha. The highlighting function you had for us to check our answers was genius as well. It's really interesting to hear that in all of the different ending versions it starts somewhat similarly and in the same key. I'm really curious now as to how you made the AI such fine details so well? Also, in terms of feedback, I noticed that your ending version that you composed seemed to have more ornaments so to speak, and sounded very elaborate which I'm sure, is not easy to do. Once again, I'm very impressed!
      Edward W
      Hi Dain! Thank you so much for all of your comments; I appreciate it a ton! As for your question, any details you hear from the AI were generated by the AI based off of the seed it was processing. So, if there are any changes in dynamics, ornamentation, or articulation, that was all put in by the AI itself. I can't really speak to how the AI was made to include these details, since I'm basing all of the AI work off of the pre-trained Allegro Music Transformer, but I can only assume that the MIDI files that were used to train Allegro Music Transformer likely included a lot of the dynamics, ornamentation, or articulation from their original pieces so that the AI could use these in its own generated pieces. I hope that answers your question, and thank you again for all your support!
    Dr. Winslow
    I have a question . . . why are there 5 audio clips for each category - they all sound the same. You take an original audio recording (the seed), and try to copy it, then give it to the algorithm to try to copy? I know I'm not understanding something somewhere with the process that you're doing.
      Edward W
      Hello Dr. Winslow! Originally, when I was using Allegro Music Transformer, the neural network would output both the original seed followed by what it would generate, so all of the beginnings sounded the same since they were all the original seed, but the endings were all different from one another. I have edited each audio file to contain just the continuations instead of the seed + the continuation now, so hopefully that confusion does not end up happening again. For your second question, the seed was used as the basis for both the AI generated music and my own continuations, so the seed was used as an input to the neural network, and the best two batches were selected to be used as the audio files in this situation. Then, I took the seed, and I composed my own continuation based on what I thought could come next. I hope that answers your question!
    isabelle_s
    Hey Eddie! Super cool and very impressive update! What kind of effect do you think that different sounding instruments might have on the listener?
      Edward W
      Hi Izzy! Sorry for the late reply. I touch on this a bit in this week's post, but the main problem I was concerned about with different sounding instruments was that they might introduce random noise into the resulting data I get from the survey. Because I want to judge the AI based solely on its compositional competency, it's important that all of the instruments use the same sound font so that a more "robotic" instrument sound doesn't influence how people respond to the survey. Thank you again for your question!

Leave a Reply