I was in a power pop sorta mood the other day, and composed this one.
Peace (pops),
Kevin
I got invited into the AI Test Kitchen by Google to begin beta testing out some early versions of their AI apps. The only one I saw available to me at this point in time was MusicLM, which was fine since I am curious about how text might be transformed into music by AI. (I’ve done some various explorations around AI and music lately. See here and here).
MusicLM was simple to use — write a text describing a kind of music (instrument, style, etc.) and you can add things like a mood or atmosphere and it kicks out two sample tracks, with an invitation to choose the best one. This is a trial version of the app and testing platform, so Google is learning from people like me using it. I suspect it may eventually be of use to video makers seeking short musical interlude snippets (but I worry it will put musicians and composers out of work).
I tried out a few prompts. Some were fine, capturing something close to what I might have expected from an AI sound generator. Some were pretty bad, choppy to the point you could almost hear the music samples being stitched together to make the file. Like I said, it’s learning.
The site does let you download your file, so I grabbed a file and took a screenshot and created the media piece above (here is direct link). My prompt here was: “Electronic keys over minor chords.” (An earlier prompt — a solo saxophone — gave me a pretty strange mix and I think I heard some Charlie Parker in there.
Here is what the Google folks write about what they are up to with MusicLM:
We introduce MusicLM, a model generating high-fidelity music from text descriptions such as “a calming violin melody backed by a distorted guitar riff”. MusicLM casts the process of conditional music generation as a hierarchical sequence-to-sequence modeling task, and it generates music at 24 kHz that remains consistent over several minutes. Our experiments show that MusicLM outperforms previous systems both in audio quality and adherence to the text description. Moreover, we demonstrate that MusicLM can be conditioned on both text and a melody in that it can transform whistled and hummed melodies according to the style described in a text caption.
I guess Google will be adding new AI-engined apps into the kitchen for testing. I’ll be curious.
Peace (and Sound),
Kevin
I’ve had time on my hands this past week and so I’ve wandered into making a few tracks of music. Here are two songs from yesterday, each capturing a little different emotional spin from being stuck temporarily at home. (Note: I think the songs are best experienced in headphones.)
Piano Crates (link)
Imagination Rhythm (link)
Peace (Rhythm and Sound),
Kevin
This post is really just an attempt of mine to gather together some of the explorations I have been doing to see what progress is being made with AI technology and the making of music and sound. It’s all pretty strange worlds out there right now.
Some sites are starting to use inputted text from users to generate sound and music. Others are built where the user does not have agency to create music, only to experience songs based on some choices, like mood or genre or artist. None of it, to my ears, sounds authentically human (yet?).
Here are a few sites I’ve explored:
Riffusion takes a phrase of writing put into its engine and moves it into a sort of song pattern, and the site features a rolling musical pattern that is designed to make visual the algorithmic patterns being used. Here is one that I did for a Daily Create that featured Riffusion with the phrase: A haiku is a moment in time. See the details on how Riffusion works — it is pretty fascinating. (Free to use)
Google is developing its own text-to-music AI generator, called MusicLM, which takes inputted text and creates a moody, evocative soundtrack to go with the words. There are plenty of intriguing examples, although it seems like Google is maybe working to figure out the copyright logistics of its database, where it has collected sound and loops that its machine uses to generate the soundtracks from text. Google also has the Magenta Project, another AI development that’s been around for bit, and the Magenta site does feature some examples of how it is being used to merge creativity with machine learning. (MusicLM not yet publicly available for use other than samples – Magenta data can be used for projects, I think)
OpenAI — the group behind ChatGPT — has Jukebox on its drawing board, and the algorithms are being fed music and sound and song lyrics, and it is learning how to create music tracks in the styles of those artists. It’s a bit jarring, to me, to hear how the machine uses Ella Fitzgerald as its source. OpenAI also has something called MuseNet, which seems similar to Jukebox. (Not yet publicly available other than samples)
The Bored Humans website has an AI Music Generator that uses neural network learning to produce entirely computer-generated songs, with lyrics. None of it is much fun to listen to for any extended period of time, in my opinion, but that it is being done is something to take note of, and worth playing around with. They even host a Lyric Generator. (Free to use)
Soundraw is a site that generates music by mood, and gives you a handful of song options to choose from. I found many of the tracks sounded the same, but I didn’t do a full explore of the possibilities there. Here is a track for the mood of “hopeful” as an example. (Free to use, but requires account for any download)
Melobytes is another odd one, but it gives you plenty of options to manipulate the sounds the AI generates as a sort of “song” from text — although every time I have used Melobytes, the song sounds pretty jarring and off-kilter to me. (Account required).
And I am sure there are others out there, too, and more to come, no doubt.
Whether this is all good for the field of music creation or not will be debated for a long time, I am sure. Just as ChatGPT and its field of chatbots has many thinking deeply on the art and creative act of writing, so too will the field of AI music generators have musicians wondering about how the field of sound is being transformed by algorithms, and what it means to be an artist in the field of music. (I also worry these AIs will put a lot of musicians who work on films and television and other media venues out of work, just as the DJ put many live bands out of work for weddings and other gigs.)
Peace (and sound),
Kevin
Music Machine flickr photo by Dogtrax shared under a Creative Commons (BY-SA) license
My connected friend, Maha Bali, shared a post about thinking through how to navigate the world of citation in the age of ChatGPT and AI-generated text, particularly if a writer uses the AI chat for parts of their own writing and/or research. Maha suggested a citation format, but I was particularly intrigued by Alan Levine’s thoughtful response in the comments, and when Alan referenced Jonathan Portiz’ insights about how or whether to reference machines in the act of writing, using the use of music software for songwriting as an example, something perked up for me.
(See Maha’s post and the comments at the bottom and then her follow-up post)
I like to write and produce music, mostly at the hobby level. Although I do play real instruments (saxophone and rhythm guitar, and I also dabble with bass and keyboards), I also often turn to apps and platforms like Garageband and Soundtrap, and use sound loops and other elements of the computer to create music.
When I have shared those pieces out, I have often wrestled with how to make sure anyone listening (if anyone is even listening) would know it wasn’t me playing those instruments, but some musical loops. Often, of course, it’s obvious, mostly because the music comes out rather way too flawless and always exactly on the beat, like a droning metronome. That said, it’s not always obvious that technology has been used. If I am layering in my own singing voice, or my saxophone, or guitar into the mix, then the hybrid pieces are a bit of both things — the human musician and the algorithmic loops.
I have yet to come to a suitable system for letting anyone listening know that a piece of music is more computer loop than musician me. To be honest, I often travel the lazy route — no mentions of the software.
Here’s an example of what I mean. A music friend had sent me some lyrics and asked for a song, which I then built musically in Garageband after adding some lyrics to his words myself, so it’s a human-human-machine collaboration. When I shared the final version with him, he admired my guitar playing, to which I let him know the reality – none of it was me.
So this topic of leaning on the machine for creativity, and whether to make that kind of technical support more visible to others in any published content through citations or some other methods, has long been at the back of my mind.
This has been made more pertinent in recent years as my teenage son has been producing his own music tracks using another (more advanced) digital music software platform, collaborating with hiphop singers and writers from around the world. He doesn’t play an instrument. He plays the platform. He doesn’t cite the platform when he posts his music on the major streaming services.
Should he be considered a musician, even though he didn’t make any of the original loops himself? What about if he edits and changes the loops, as he does? Should every loop he uses be cited somehow?
All this brings us to ChatGPT and its brethren, and Maha’s exploration of how to make citations when using AI chat platforms for writing pieces.
Is it necessary to cite the machine?
My initial impulse is that Maha’s discussion about writing and citation feels different from making songs because it is writing of words through predictive text of the AI and not music composition with prerecorded loops. Writing a poem or a story or an essay also feels different than writing a song that layers words over music.
Even as I write that, though, I realize: that statement doesn’t seem to sit well with me at all — all are creative acts that begin with nothing but an idea and lead to something that others can experience. Maybe my conflicted feelings stem from being so used to technology being integrated so fully into the modern field of music production, and I am not yet used to its use in the field of writing.
Not yet, anyway. Will time and experience change that?
Garageband and Soundtrap and others don’t cite the musicians where the original loops came from. Do we expect that ChatGPT and others will cite where their words come scraped from? I believe that to be a strong yes in my view – that such information about original sources should be baked into the chart system (even as I understand the technical aspects will make such a thing nearly impossible). If this were done, then a writer could cite the sources of their AI-influenced writing.
Right? Hmm.
How confusing is all this? Pretty.
And how interesting to grapple with it? Very much so.
Peace (in the machine),
Kevin
I was working on a song the other day, started early after a restless night (not all that unusual for me), and needed a title. I came up with Slow Walking Through Sleep, and as I was pulling the music beneath an image, words to a poem began to filter in. So I added that, too. I debated about whether I should add my voice, reading the poem, but opted against it, to allow the music that started it all to shimmer through.
Peace (and Song),
Kevin