#341 – Making Music With Generative AI

#341 — Making Music with Generative AI

AVID Open Access
12 min

Keywords
generative AI, music production, text to music, Google Music FX, Suno tool, AI-generated music, song prompts, instrumental mode, vocal generation, AI-generated lyrics, podcast music, AI impact, classroom use, AI ethics, AI-generated content

Transcript

Paul Beckermann 0:01
Welcome to Tech Talk for Teachers. I’m your host, Paul Beckermann.

Transition Music 0:06
Check it out. Check it out. Check it out. Check it out. What’s in the toolkit? What is in the toolkit? So, what’s in the toolkit? Check it out.

Paul Beckermann 0:16
The topic of today’s episode is Making Music with Generative AI. Generative AI continues to evolve. It’s gone from text to images to video and audio. Of those four types of AI-generated media, however, I feel like audio has been a bit overlooked, and specifically, music production. As someone who writes, performs, and records his own music, this one got me curious. So in today’s episode, I’m going to give you a quick introduction to text-to-music AI generators. There are actually quite a few of these tools available online. As with anything AI, the sheer volume of options can feel overwhelming, so I’m going to focus on two tools as examples of what’s out there: Google’s MusicFX and a freemium product called Suno. Before I dive into the tools, let me give you the big picture. AI music generators work much like text-to-text or text-to-image generators. The AI is trained on existing examples, in this case, music, and it learns how to generate new music based on these models. The AI learns things such as instrumentation, musical style, song structure, and lyric composition. It then takes what it has learned and applies it to the unique prompts that you, the end user, enter. If you ask for a song in the styles of 1980s hard rock, the tool will review the songs it’s been trained on that match that descriptor, and then it will generate a new song based on the elements of those sources. It’s a bit like asking a human musician to play a riff on a guitar that sounds something like he or she would have heard in the ’80s. That musician would likely respond by playing something based on personal memories and experiences playing covers of ’80’s songs or examples heard on the radio. The musician would play something new but similar based on those past examples from that time period. That’s what the AI is doing. With that context in mind, let’s explore some examples.

Transition Music 2:22
Here’s the, here’s the, here’s the tool for today. Here’s the tool for today.

Paul Beckermann 2:27
The first example I want to share is a free one from Google Labs. It’s still under development, but it’s free and offers an entry point into text-to-music AI generation. It’s called MusicFX. To access it, you’ll need to navigate to Aitestkitchen.withgoogle.com/tools/music-fx. Now, I know that’s a long address, so I’ll put the link in the show notes to this episode on AvidOpenAccess.org, but you should be able to Google it and find it, as well. Once you’re at the site, you’ll need to set up an account or log in with your Google credentials, which is what I did. You’ll be greeted by a fairly simple screen with a text box where you can type in a description of the music you’d like to be generated. Google provides a few tips and a virtual tour of the page the first time you log in. For example, it suggests: “To get started, type in a simple description of the song you’d like to create. For best results, include descriptive genres and moods.” There are some additional quick links at the bottom of the screen that you can click, terms such as soothing, jazz, funk, and saxophone. To try this out, I entered the following prompt: “a positive, upbeat ’80s, guitar-heavy rock song about graduating from high school.” From the simple menu on the screen, I could choose the track length from 30 seconds to 70 seconds, and I could choose whether or not to have the audio loop. Looping means that the end and the beginnings of the track run seamlessly into each other, allowing it to be played over and over for an extended music track. I could download the track as an MP3 or get a shareable link. Here’s a portion of the track that was generated.

AI-Generated Song 4:08
[AI-generated instrumental music plays.]

Paul Beckermann 4:27
Pretty impressive, huh? All that from a simple prompt. One interesting feature is that once you’ve generated the music, your prompt becomes interactive with keywords from the prompt turning into drop-down menus that offer suggestions for modifying that prompt. For instance, in my prompt, the word “rock” turned into a drop-down menu that allowed me to pick rock, pop, country, or jazz. Similarly, the words heavy-guitar became a drop-down that also gave me options for synth heavy, drum heavy, and vocal heavy. I felt like it offered a way to rethink the prompt and come up with new iterations that maybe I wouldn’t have thought of otherwise. Each time I clicked “Generate Music,” I got two versions of the song to choose from, although I did find that the two options were usually pretty similar. This tool from Google only gave me instrumentals. Even when I instructed it to include specific lyrics in the prompt, it only gave me music. There might be a way to get lyrics and vocals, but I didn’t see one.

The second tool I tried was a freemium product called Suno. With the free version, you can generate about ten 3-minute songs per day. With the premium version, you get more credits to produce more songs, longer songs, and you also purchase ownership of the track so you can use it commercially. I tried both the free and premium versions. The premium version cost me $10 for a month, and allowed me to generate up to 500 songs for that month. Suno works very much like MusicFX, but Suno offers more functionality, the ability to produce vocals and, what I felt, were much more impressive results. For my experiment, I again entered the same prompt: a positive, upbeat ’80s, guitar-heavy rock song about graduating from high school. When I clicked “Generate,” I quickly got two versions of full songs to choose from. Both versions included music, vocals, and lyrics to a song titled “Highway to the Future.” I could have provided the lyrics myself if I wished, but I decided to let AI do the work for me. The lyrics were the same for both versions. Here’s a bit of version one.

AI-Generated Song 6:48
Caps in the air. Dreams are flying high. No more halls, no more reasons to lie. Freedom calls, the horizon wide and bright. Graduation day, world within our sight…

Paul Beckermann 7:03
Here’s version two, which I think I like better.

AI-Generated Song 7:16
Caps in the air. Dreams are flying high. No more halls, no more reasons to lie. Freedom calls, the horizon wide and bright. Graduation day, world within our sight. Speeding down that highway, Destiny. Classrooms…

Paul Beckermann 7:39
As you can hear, this music is of higher quality than MusicFX. It’s quite a bit more sophisticated as a composition and includes fairly realistic sounding vocals. It’s also longer and is structured like a real song with verses, a chorus, and a bridge section. I could download these files as MP3 with the free version and higher quality WAV files with the paid version. After reviewing the lyrics and music from the first generation, I decided to modify my prompt slightly to see how much the tracks would change. I edited the prompt by adding the following phrase to the end: “Avoid trite or overused phrases.” Interestingly, I didn’t think the lyrics were any better on the new version, but the music was heavier. Here’s what I got.

That is a lot different result from just a few different words. After playing with Suno for a while, I began wondering how this tool might work for generating audio to use in a podcast. So I entered: “Catchy but subtle techno music to open a podcast about K-12 education,” and then I turned on the instrumental mode. Here’s what I got.

AI-Generated Song 9:22
[AI-generated instrumental music plays.]

Paul Beckermann 9:30
I thought that was pretty good. In fact, all of these results were fairly impressive, and because the paid version is so inexpensive and so easy to produce tracks, streaming services are beginning to get flooded with this type of AI-generated music. It makes me wonder what Spotify is going to look like in a few years. As a musician, I’m not sure I love this. I’m still a fan of imperfect human compositions. Still, I’m impressed, and I’m a realist, and I know that these types of tools aren’t going away and will continue to evolve. And I do think they can have usefulness. Here are a few ways that I might consider using this in my classroom.

Transition Music 9:30
How do I use it? Integration inspiration. Integration ideas.

Paul Beckermann 9:32
Number one, create audio for a podcast or video soundtrack where the music is more for texture than the main event. Number two, analysis of AI generation. I could see having students compare the different outputs based on varied prompts to explore the cause and effect of these changes. That could lead to some really great analysis and discussion. It would also teach them prompt engineering. And number three, explore the idea of AI music generation. You could play these types of clips for your students, demonstrate how they’re generated or do it live in class, and foster discussion about the impact of this type of AI. Again, this could lead to a rich discussion and give students a voice in talking about the ethics of AI and its potential impact in the future. They could share how they feel about it, as well as when they think it’s appropriate, and perhaps inappropriate, to use this type of tool. You might even have them produce a best practices user’s guide for using music produced by generative AI. So there you have it. A quick introduction to AI text-to-music generation. It’s one more way that AI is impacting the creation of new content. To send you on your way today, let me leave you with one more tune. This one was generated with the prompt; “Create a beat-heavy and catchy jingle for a podcast called “Tech Talk for Teachers.” The only lyrics should be the title and this can be short.” All right, so how about this? Here you go.

AI-Generated Song 11:47
[AI-generated beat plays.]

Paul Beckermann 11:47
To learn more about today’s topic and explore other free resources, visit AvidOpenAccess.org. Specifically, I encourage you to check out the article collection, AI in the K-12 Classroom. And, of course, be sure to join Rena, Winston, and me every Wednesday for our full-length podcast, Unpacking Education, where we’re joined by exceptional guests and explore education topics that are important to you. Thanks for listening, take care and thanks for all you do. You make a difference.