So What? AI Music Generation for Marketing - Trust Insights Marketing Analytics Consulting

So What? Marketing Analytics and Insights Live

airs every Thursday at 1 pm EST.

You can watch on YouTube Live. Be sure to subscribe and follow so you never miss an episode!

In this episode of So What? The Trust Insights weekly livestream, you’ll learn how AI music generation tools, like Suno, are changing the landscape of music creation. You’ll discover how these tools work and the ethical implications of using AI-generated music. You’ll also learn how to use these tools effectively and ethically in your own work. Tune in to discover how to leverage AI music for your marketing needs.

Watch the video here:

So What? AI Music Generation for Marketing

Watch this video on YouTube

Can’t see anything? Watch it on YouTube here.

In this episode you’ll learn:

Use cases for AI music generation
Ethics of AI music generation
A proven process for improving AI music generation with tools like Suno

Transcript:

What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.

Christopher Penn 00:34
Well, hi, everyone. It’s Thursday. This is the marketing, analytics, and insights live show—although I think we need to also add “AI” in there somewhere because that’s been our thing for the last few months, slash, years. John and I are here today talking about music and AI while Katie is off hunting down bears with a camera in northern Canada. So, John, how are you?

John Wall 00:59
Good, good. Yeah. Peak foliage for her up there. That’s good. I’ve been entertaining myself with New England’s favorite fall pastime: heating system roulette as we fire up the heater and see what happens. So it’s, yeah, all fun.

Christopher Penn 01:12
Oh, that’s delightful. Today, we’re going to be talking about music generation and AI, some of the use cases for this, touching on the ethics of it, and then the process for how to do it, particularly with tools like Suno. So, John to start off generation of AI music, today’s tools like Suno and udeo and Music Gen and stuff can create pretty coherent music. It’s not going to win any Grammys, but it also does not sound like you’re face-rolling on a piano either. When you think about the use cases for audio, for music in particular, what are the ones that really spring to mind for you?

John Wall 01:55
Yeah, there’s a couple different levels. I mean, it’s very similar to what we saw with generative AI images. The big one is for most of us and our watchers and listeners: you’ve got something, and you want to put some music there. It does not need to be anything great. It just needs to be something basically as a background track to hide all the noise and all the screw-ups, something behind there. And, you know, if it is closer to the mood of what you want to do—like, that’s fantastic, too—but you just need something good enough to get there. But then you’ve been playing around a lot with this stuff, and I’m interested in some of the stuff where it’s actually starting to sound pretty good.

John Wall 02:32
Like it’s stuff that you would leave on in the background as you work and not think that this was just all created by a computer. So, yeah, there’s going to be some interesting stuff about leveling the playing field for everyone to be able to create music and kind of get their artistic stuff out there.

Christopher Penn 02:49
And that’s a really, I think, critical point is the content that these things create today certainly is good enough for, I would call it, commodity music. So, hold music is an example. There’s no reason why you can’t have just a bunch of hold music for elevator music for things where music is not the central focus of what you’re doing, as opposed to sitting down with your, your Sennheiser, your studio headphones and actually listening to it. It’s probably not going to be something you’re like, “This is my music listening session.” You’re not going to do that because it’s not as good as human-led music.

John Wall 03:32
So, how about what’s even in the tool stack? I haven’t even been watching this space because I’ve done a bunch of stuff with images, but have not gone anywhere near music.

Christopher Penn 03:40
So, the two big players right now that have commercial offerings, there’s one called Suno and one called udeo. Suno, which is at suno.com and then Udeo, which is obviously at udeo.com. These are the two big ones of the two services. Suno is kind of the market leader right now for a couple of reasons. One, it can generate tracks up to four minutes long. Udeo is still limited to roughly about 30-second segments. Interestingly, behind the scenes, if you look at the structure of the music that Suno produces, it produces it more or less in 30-second chunks that then silently seem to glue together because you can have some hallucinations—some auditory hallucinations—in the background. And then there are the big libraries like MusicGen, which I believe is Meta’s, and Google’s got their own music library as well. But, those are not commercially available.

Christopher Penn 04:43
Those are research tools, and you can use them, but their quality is not great. What the teams at the commercial companies are doing seems to be good enough, again, for basic stuff. So, to give you a sense of some of the output, this is just a very light jazz piece that the machine put together from a pretty straightforward prompt. So, let’s give, like, a few seconds. Listen to this. Now, I was speaking at an event not too long ago and—one of the, it was interesting—one of the things that the organizer said was, “Hey, we don’t like to have just, like, dead space for the 45 minutes that people arrive and have, drinks and whatever. Do you have a playlist that you can play?”

Christopher Penn 05:39
And I said, “Do you have an ASCAP or BMI license, which a venue is required to have if they’re going to play copyrighted music? If you do not do it, a music industry lawyer shows up and punches you in the face, and the fines are substantial.” AI-generated music falls into the same category as AI-generated anything. It has no copyright, and therefore, it has no legal standing, but it also has no licensing either because it was not made by people. And therefore, if you play AI-generated music and, say, a venue of some kind, you don’t need to have those licenses. So, for, in this case, the music was not the centerpiece, it was just background. And, so this was a perfectly fine use case, which brings us to the ethics of it. So the ethics of this.

Christopher Penn 06:31
And, there’s problems in music generation. Two of them, two big ones. Number one, in the example I just mentioned, when I play AI-generated music, no one’s getting paid. There is no artist, there is no musician, and therefore, any money that a human artist would make isn’t happening. Interestingly, when you look at some of the major music services like Spotify, they do nothing, exclude AI-generated music. And, there is a belief—I don’t have any documentation to back this up, but there’s a belief, particularly in the music community—that these services will embrace AI-generated music because it means they don’t have to pay musicians for the same reason that Spotify embraced podcasts.

Christopher Penn 07:15
Every minute that you’re listening to a podcast on Spotify is a minute you’re not listening to licensed music, which means it’s coins that Spotify does not have to pay out to anyone.

John Wall 07:26
Yeah, it’s definitely the impact on the. This is just classic “digital eats everything.” So, yeah, there’s a lot of life for the independent artists is not going to get any better anytime soon.

Christopher Penn 07:38
Yep. The other side is the training data. The companies, the commercial companies, do not disclose what training data they used, but it’s pretty apparent that they probably scraped YouTube as most companies do. And, the reason I say this is that, in these tools, they have some basic copyright protection. So, if you were to copy and paste Taylor Swift’s lyrics in, it will say, “I’m sorry, this is a copyrighted set of lyrics. You can’t use this.” If you alter enough that it is semantic, linguistically different—but it’s semantically similar—so you take the lyrics, change, like, 10 or 15% of the words, enough to break the copy protection filter, it will ingest it, and then it will return a song where chunks of it sound eerily like a Taylor Swift song that you might recognize.

Christopher Penn 08:32
We’re not going to do this because I don’t want Taylor Swift’s lawyers to.

John Wall 08:37
Come to your door.

Christopher Penn 08:38
Yeah, exactly. But, my friend Ruby King and I were messing around at Macon at the marketing AI conference with a few different scenarios, so. And, I absolutely will never publish this information, ever. She had me sing a few different songs using the musical notes and tunes of other songs. So, I was asked, and I did sing “Call Me Maybe”, but to the tune of “Happy Birthday To You”. And then, we fed that into the new cover feature and said, “Cover the song,” and chunks of it were the exact melody of “Happy Birthday To You”, and chunks of it, because of the language used, were “Call Me Maybe”. And, there was the chord progression, you could tell that pretty clearly came from copyrighted training data.

Christopher Penn 09:34
So, the caution here with these tools is that they are, for sure, using copyrighted data. And if there’s one industry that’s really good at suing people, it’s the music industry. So, I know Suno itself has some pending lawsuits about the use of copyrighted data, so just be aware that the tools are not without their ethical complications.

John Wall 10:00
Yeah. And, what do you see? I don’t know. This is. I mean, we’re seeing this all over the place. I mean, these large language models have no life experience, so really, everything in there is taken from some other artwork. The problem is we can’t attribute any of it back to where it goes.

Christopher Penn 10:15
Right?

John Wall 10:15
It could be a penny for everybody, or there could be select artists that do a bigger chunk, but we don’t know anything about that. And, that’s not even getting into the thorny mess of, these companies are startups. It’s not like they have $20 billion lying around to just start paying out to artists. So, there’s all of that. But, I don’t know, do you see any path for where you think this is going to go over the next few years?

Christopher Penn 10:37
I do. And, you’re not going to like the answer. If you’re a musician, the answer is, these independent startups probably are going to get bought, or they will be successfully sued because, particularly with music, it’s pretty clear. Like, when you’re. When you could reproduce the melody and the harmonies of “Gin and Juice”… Yes, Snoop’s coming for you. However, there is an exception. There’s a loophole. If you have the rights to use someone’s work, then you no longer have to.

John Wall 11:12
You.

Christopher Penn 11:12
You have a license for it. Who has those rights? Well, who owns YouTube? Google. And, when you upload your content to YouTube, what do you sign away? You sign away all of your rights to your content. You say if you load your content to YouTube, you give Google the rights to do pretty much whatever it wants with it, including create AI models from it. So, Meta, same thing. If you are a musician, you post your song on Instagram or whatever, guess what? You have also given Meta a license to use that for whatever they want. And, you can’t take that back once you’ve agreed to the terms of service.

Christopher Penn 11:50
So, I have absolutely no doubt that within the walls of Meta and Google, they have working music generation models that are top notch because they own the data that we’ve all been providing them for 20 years now.

John Wall 12:06
Free ride. Yeah, that makes sense.

Christopher Penn 12:10
Which means that if they do, at some point, release that or release tools to do that, the music industry won’t have a leg to stand on because all these artists that have loaded their music and videos and stuff to YouTube and to Meta’s properties over the last two decades have also signed the license to say you can do whatever you want with my content. So, that’s the ethics portion. I think it’s important to cover that. So, how do we use these tools—Suno in particular? Well, I’ll walk through the basic process. There’s a limited number of credits you can use free, or you can pay—I think it’s, like, $90 a year to use their service—and you get a certain number of credits per year, per month. And, then each song costs a certain number of credits. It’s, like, carnival tokens, almost convoluted.

Christopher Penn 13:00
You load—if you’re doing a song with lyrics, you can, well, you can do it very simply—you can give it a very short 200-character prompt and it will try to create something or you can go into custom mode, choose the model you want to use—and generally speaking you want the newest, best model. You provide your lyrics and your style of music and then a song title, and then it goes and it begins to generate. So, here’s the thing: their lyrics generation software is terrible because it’s just not a super high-quality language model. Whereas, if you use a big language model, it does a pretty good job. The style of music stuff, they tell you the sort of style, but it also takes in other things.

Christopher Penn 13:47
The thing that people don’t know about this tool in particular is the lyrics are part of the prompt for the music. So, depending on how the lyrics are structured will condition how the model creates.

John Wall 13:59
Its music as far as everything from, whether it’s a genre or just kind of how the song ends up getting laid out as far as tempo and where the brain is and that kind of stuff. Both. All that stuff.

Christopher Penn 14:14
Yeah. So, it’s controlled by all that. So, to get the most out of this tool, you probably want to do your lyrics and your style of music prompts in a foundation model—a big model like ChatGPT or Google Gemini or Anthropic—and then copy and paste and port it into Suno to get the best results. At least that’s been my experience with it. So, let’s go ahead and do that. We’re going to start just with an example. We can take existing lyrics and move them in. So, this, our friend Brooke Selles posted. She had ChatGPT write a song about a survey she’s trying to have people do. So, we can take this structure of lyrics here. And, then one of the things we need to do, we need to remove those emoji because those emoji are not going to be our friend.

Christopher Penn 15:07
And, I will take that, these lyrics here, and we’ll put this straight into Suno’s lyrics section. Now, that’s about 1,200 out of 3,000 characters. That’s, so that’s well within the rates. The next part is the style of music, and this is where we want to provide really good feedback, some technical stuff. So John, for a song that our friend Brooke has written, what would you say would be a good music style for that?

John Wall 15:35
Oh, that’s got to be pop music. Yeah, of course.

Christopher Penn 15:38
Okay, pick an artist.

John Wall 15:41
Oh, let’s see, who would we want it to be?

Christopher Penn 15:45
Or pick a song. Pick an individual pop song that you can think of.

John Wall 15:48
I’m trying to think of. How about “Houdini”?

Christopher Penn 15:55
Who’s that by?

John Wall 15:56
That’s. Let me make sure I get this straight because there’s, like, five artists that, to me, it’s like they all sound the same. Dua Lipa, das Houdini.

Christopher Penn 16:07
Okay, so first, go to ask Gemini, “What do you know about this song?” And, it’s going to spit out a bunch of stuff, and there’s some basic musical stuff. So, we’re going to say, “Describe the technical aspects of this song such as key, tempo, melody, heart harmony, chord progression, instrumentation, etcetera.” So, we’re going to ask Gemini, what does it know about these pieces? It: key, G major; 128 beats per minute; melody, harmony, chord progression. And, we’re going to next say, “Write a prompt to give to a music composer to have them write a stylistically similar song to ‘Houdini’ incorporating these technical aspects.” So, we’ve now gone from getting its knowledge, the primed knowledge, and turning it into, “What are the tech specs of this song?” This is a useful, hefty amount of stuff, so. But, it’s too big for Suno’s prompt.

Christopher Penn 17:35
Say, “Condense this prompt down to 200 characters, preserving the most technical aspects, most key technical aspects. Aspects omit verbs.” So, it should now turn this into a Suno-style, essentially a Suno-style prompt. So, there’s that. We can’t use the copyrighted terms. It will, it will flag on that. So, let’s take that. Paste that in here. See up.

John Wall 18:07
Sorry.

Christopher Penn 18:07
It’s 120 characters. I forgot about that.

John Wall 18:09
Oh, wow. Less than a tweet.

Christopher Penn 18:11
Condense down to 120 characters. I’ve got 200. If you don’t provide lyrics. Okay, so G major, 128 beats per minute. Diatonic new disco. All right. And then, we’re going to call this “Take My Survey, Please”. Number one. What happens next is that it will evaluate the prompt. It will evaluate the style of music. You can tell it to exclude certain styles. Like, if you know that it has a tendency to spit out a certain type of prompt or a certain type of song, you can exclude, like, “Please stop making this sound like, country pop.” And, you can say, “I don’t want it to sound like country pop.” And then, once that’s done, you can listen to a preview right away, or you can wait for it to, the MP3 to download.

Christopher Penn 19:12
We can see it’s pending right now, but that’s the general process for doing this if you have lyrics. If you don’t have lyrics, you’d want to repeat a similar process, but with the construction of lyrics as well, which is considerably more challenging. What do you think so far?

John Wall 19:30
Yeah, the thing that shocked me the most was in the other model where you said, “Break down the song,” and it gave you the full profile of the thing. Like, that’s kind of crazy. I was not expecting it to be that in depth and on the mark. And, then the other thing. So, here it’s spit out, too, is that just giving you two variants to check out. So, you’ve got two flavors, and you would continue going in one direction.

Christopher Penn 19:49
Exactly. So, if we could get the audio hookup working on this, you know what the heck? I’m just going to go for it. And, I can’t stream off of my—this tab—but I am going to try and stream from my phone. So, let’s see if we can get Suno to play one of these candidates. Let me know if you hear anything.

John Wall 20:22
Yeah, it’s coming through.

Christopher Penn 20:23
Really? Because I can’t hear it at all.

John Wall 20:25
You can’t hear it?

Christopher Penn 20:26
No.

John Wall 20:26
This is the joy of computer audio.

Christopher Penn 20:31
Is it any good?

John Wall 20:36
It’s pretty good, yeah.

Christopher Penn 20:39
Okay, I’m gonna pause it. Oh, that’s delightful. Okay, now to do this with lyrics again, you probably don’t want to use Suno’s built-in stuff. Instead, what you’d want to do is build a lyric construction framework. So, let’s go ahead and do that. I’m going to use Google’s AI studio, which is the more advanced version of Gemini. And, I’m using this because you can control which model to use. I’m going to use Gemini Pro zero two, which is their latest, greatest, fanciest model. I’m going to change the output length to 16,300 tokens, which is about 12,000 words, which means that it can foam with the mouse pretty significantly. What kind of song do we want to create? Pick something a bit more challenging.

John Wall 21:34
How about a seventies funk like Earth, Wind & Fire? Something like that.

Christopher Penn 21:38
Okay, so the lyrics for that. So, all right, we’re going to start by saying, “What do you know about the writing of song lyrics in the style of 1970s funk? For example, Earth, Wind & Fire lyrics?” So, we’re going to ask the model, what do you know about this? While it’s spitting that out, we’re going to start the next prompt saying, “What are some common lyrics writing mistakes that less experienced songwriters make when writing 1970s funk lyrics?” So, this three-step process is part of the Trust Insights pair framework, which you can download for free with no information required at Trust Insights AI pair. So, there’s our less experienced stuff, and then we’re going to say, “What expert tips and tricks do you know about writing 1970s funk music song lyrics that we have not discussed yet?”

Christopher Penn 23:02
So, this is now digging—forcing the model to basically say, “Here’s what we haven’t thought of yet.” So, we’re trying to broaden the field of knowledge within this conversation. Once we’ve done this, we want to ask it, “Create a consolidated, de-duplicated, comprehensive list of best practices for writing 1970s funk song lyrics based on everything we’ve talked about so far.” This step, in addition to consolidating the information, also reinforces the in-context memory. So, when you’re using language models, every part of what’s happened in the conversation already becomes a part of the next prompt. So, all the history becomes part of the next prompt. So, by having it reiterate everything you’ve already done, you’re essentially increasing the weight of specific, the specific words and phrases that are most relevant by having it summarized.

Christopher Penn 24:10
And also, for smaller models in the old days, this was also a good technique to use to refract its memory. It’s not as concern anymore with today’s models. So, this is a pretty good framework for writing seventies funk lyrics. Next, what we’re going to do is we’re going to have it build a scorecard. So, we’re going to say, “Next, build a scoring rubric to assess the quality of a set of song lyrics that are 1970s funk style. Score the song lyrics based on their adherence to the best practices. Individual items from the rubric should be scored in variable integers: zero to 30, to 50, to seven, etcetera. The rubric should total 100 points. Build the scoring rubric.” So, this scorecard is a form of self-evaluation.

Christopher Penn 25:25
What we want the model to do is to come up with a way to judge whether its own generation or any sort of song lyrics would conform well to these best practices. We can see it’s got things like musicality, structure, dynamic build, etcetera. I’m going to go ahead and copy this in markdown format. I’m going to store it to the side. What we’re going to do next is we’re going to have the model turn this into instructions. So, we’re going to say, “Convert everything we’ve talked about so far into system instructions for a large language model like Google Gemini. The purpose of the system instructions is to accept input from the user about a topic or an existing work that they want to convert into 1970s. The system instructions will accept the input and then follow these steps autonomously.”

Christopher Penn 26:49
“One, accept the input from the user. Do not ask the user for additional input. Two, generate three variations. Three candidates of song lyrics based on the user’s input, adhering to the 1970s funk lyrics best practices. Three, score all three candidates using the—” what is it called here? “—The 1970s funk lyrics scoring rubric. Four, choose the highest scoring candidate and make an additional refinement of it designed to increase its score in the five, score the refinement. Six, produce the highest scoring candidate of the song lyrics in full. Write the system instructions in the second person imperative.” Okay, so what we’re doing here is we are building essentially a botanical, an application that will create song lyrics that will match this. And what this is going to do is this will instruct the language model about what to do. So, let’s go ahead and copy this.

Christopher Penn 28:44
Put it in a text document like so, and let’s take a look here. So, we’re going to need to put the scoring rubric into this prompt. So, we’ll denote the space. So now, and let’s just make sure that we have, we don’t have too many extra spaces. Good. All right, this is the prompt. Now, what do we do with this thing? There’s a bunch of different things we can do, but for today, let’s go ahead and start a new prompt, and we’re going to paste the system instructions in right in here. That whole big honkin’ thing. Now, let’s take Brook’s song, and let’s convert it to 1970s funk. Here are the lyrics to convert to 1970s funk. Now, let’s watch it work.

John Wall 29:48
Get down with the feedback.

Christopher Penn 29:54
Interesting that it actually skipped ahead and went with sending it. Let’s, I don’t think it. I don’t think it did a great job with that. So, let’s. Let’s delete that. Well, let’s revise the prompt. “Follow the system instructions for the 1970s funk lyric generator strictly.” Interesting. It is not, it does not want to step through the instructions. That’s okay. Wow. Oh, this is fabulous. Let’s go ahead and actually, let’s try making this into a gem. Let’s see if a gem from Gemini works any better at that. Let’s go ahead and start this seventies funk.

John Wall 30:54
Oh, well, you’re hammering at that, too. I noticed that there’s no way to take a recent and make it into a gem. You have to just make a gem to start.

Christopher Penn 31:02
Yes. You have to make it just to start. And, let’s try this. And, let’s take our initial prompt and put this in here and see what happens. See if this is any better. Instruction following.

John Wall 31:25
Yeah, same deal, huh? It just drives.

Christopher Penn 31:29
No, it’s actually, this is executing properly, so it’s generating candidates.

John Wall 31:33
Oh, it’s doing multiple. Oh, yeah, there we go.

Christopher Penn 31:36
Hey, mama. Something to say? “Refined candidate three. So, candidate three scored the highest final score, 90.” Okay, so candidate three is the. The final version that seems to have done. So, now, let’s take those lyrics. Let’s save that to the side because now we need to get the musicality of it. So, we’re going to repeat what we did with Dua Lipa. “What are the musical characteristics of 1970s funk, such as the works of Earth, Wind & Fire: rhythm and groove, syncopation, bass line prominence, instrumentation, and harmony.” Okay, this is good call and response. Let’s see. “Write a prompt for a music composer to compose music in the style of 1970s funk.” Okay.

John Wall 32:54
It totally pulled in Funkadelic and Kool & the Gang, too. It made that leap. That’s excellent.

Christopher Penn 33:03
Condense this prompt down to 120 characters, preserving key technical elements such as tempo, instrumentation, et cetera. Omit verbs. All right. “Funk 110 base.” Okay, we can’t use, obviously, the names, so we’ll just take this portion. Let’s go ahead now and create ourselves a new song here. Oh, it’s not gonna let me do it because I’ve already got this in the window. Change to two, and we’re gonna take those lyrics that it generated. Double-check. Oh, it had a bit of a misspelling. So, let’s create candidate three that has the correct spelling of “harmony”. Survey two is ready to start previewing. Survey three is on the way. On the way, so. Oh, there we go. All right. Are we feeling lucky?

John Wall 34:16
There’s only one way to find out.

Christopher Penn 34:19
I’m going to have to watch the replay because I can’t hear this. Let’s see. Oh, it’s not rendered yet.

John Wall 34:34
It’s still thinking.

Christopher Penn 34:36
It is still thinking about it. I’m trying to share it from my phone, and it does not want to play.

John Wall 34:52
It’s funny, if this even comes in 20% Earth, Wind & Fire, I’m going to be laughing like crazy.

Christopher Penn 34:59
That’s version one. We don’t want version one. We want version three. Let’s ditch that. Okay. Library, my songs. Yeah. It’s not, it is not rendered in the app at all.

John Wall 35:26
It’s amazing to me that this can even, both synthesize the prompt and then create an mp3 file and a sample. I mean, that’s just an insane lift.

Christopher Penn 35:47
I’m going to live dangerously here. I am going to stop sharing this, remove this one, and then I’m going to share just this, the tab. This may completely blow up this live stream.

John Wall 36:01
We’re live without a net here.

Christopher Penn 36:03
We are absolutely without a net. Let’s hear. Let’s go. It hosed the voices on that one because there were lyrics. You can’t hear them.

Speaker 3 36:43
Mama got a little something to say, sent you a message, hope you’re having a good day, slid in your DMs with a funky, “Don’t leave me hanging, let your answers glow, fill out my survey.” Yeah, it’s a soulful plea. Just a few questions, set your spirit free. Help me rise up, don’t let me fall, social care is the answer for one and all.

John Wall 37:19
Horn is totally making me laugh. That’s, like, something from a 2003 video game, but it’s got the moving bass line and the lyrics. And the vocals are straight to the point, so. And, this is usually party music. Nobody’s expecting Thoreau for these lyrics, that’s true.

Christopher Penn 37:43
I wonder what happened to version two. Let’s see if version two works at all here.

John Wall 38:00
I’m getting very… “Seinfeld”.

Christopher Penn 38:05
What the heck? Yeah. So, it hosed the vocals on that.

John Wall 38:13
That’s weird. And, then plus the garage door opening sound. Exactly. That was all about.

Christopher Penn 38:21
This is also, you know, one of the challenges with these tools is that they will create. They’re forecasting—essentially, they’re doing predictions—based on the training data they’ve seen. So, it is entirely possible there is a seventies funk song that incorporates the music of, like, thunderstorms or whatever that anomaly was that it pulled in from its, in terms of tokens that it was aware of, and that created that very interesting result that’s totally apart.

John Wall 38:49
I could totally see George Clinton, one of their 40-minute jams having all kinds of bizarro stuff in it.

Christopher Penn 38:55
Exactly. So, that’s the process that I would recommend using to generate music, particularly if you’re doing stuff with lyrics. Have a foundation model, a frontier model, like ChatGPT or Gemini or Anthropic’s Claude, build the lyrics. Have, ideally, a web-aware foundation model. Build the music composition prompt based on what you want so that you get as many of the technical terms needed in the style prompt as possible because even a single word makes a difference in terms of how the songs come out. And so, if you’re burning characters on prepositions and punctuation and all stuff, you’re not making the most of the tool to squeeze as much as you can in that space. And, remember that the lyrics themselves are part of the prompt. So, your lyrics should be structured in such a way that they.

Christopher Penn 39:53
They help the model understand what kind of music it’s creating. So, with this, John, what are you going to go create?

John Wall 40:01
The thing that I really want to see, like, I just can’t wait for the next generation, like, having to boil—because that prompt was amazing—having to boil that down to 120 characters is a crime. Like, the day you can grab that prompt and throw that whole prompt in there, that, there’s going to be some real interesting stuff. But, yeah, it just seems like there’s a lot of fruitiness going on when you’re having to crank the vice so tight on that training data, exactly.

Christopher Penn 40:26
One other thing that my friend Ruby talked about was that these tools will sit. Because they’re sampling essentially existing music, they will sometimes come up with sequences and things that if a trained, musician hears, it will know, “That was 100% machine generated,” because the physical agility needed to play what the machine came up with is impossible. She was pointing out with violins, there was one section of this one composition we created where she said, “You would have to be on, like, meth to be able to fiddle all those notes that quickly, in that order.”

John Wall 41:01
Yeah, that’s not surprising. You’re not limited at all by the limitations of the physical world and trying to do. Yeah, that’s a whole nother realm of musicality that I don’t pretend to know anything about. But, I do know that there’s, yeah, occasionally there’s songs where people are just like, “Well, how did you string your guitar to do that? Because you can’t do that with a regular guitar.”

Christopher Penn 41:19
Exactly. Or, “Hey, you must have, like, 18 fingers to play that piano sequence, or three hands,” or something. So, be aware that as with all things AI, you should be disclosing that content is AI-generated. If you don’t, ethically, that’s not a great thing to do, but, a trained musician will go, “You used AI for that. And, here’s how I know.”

John Wall 41:44
I’m just dying. This is a variation on “too many fingers.” This is the same thing we see with images. You’ve got too many fingers in the output.

Christopher Penn 41:51
Exactly. It’s just too much, too many fast hands on the keyboards and the, and the strings. In terms of when you should use AI versus when you should use a human musician, I think the answer is pretty clear. For music where it doesn’t matter—like, hold music, like, background music, like music just as a backing track on a podcast, just some percussion or whatever—I think that’s fine because, a, you probably aren’t going to use, you’re probably going to pay human musicians in the first place, and b, it’s inconsequential if the machine doesn’t do a great job. Equally true is if you’re going to do something novelty. So, as an example, probably no one would ever pay a few hundred bucks to a musician to take Brook’s lyrics and reformat them into a funk song.

John Wall 42:51
Right.

Christopher Penn 42:51
That’s, we did not take work away from a human musician.

John Wall 42:55
Right. That was never going to be a paying gig.

Christopher Penn 42:58
Exactly. For anything that you need to own the copyright on, you must have a human do it because, right now, all AI-generated music is non-copyrightable. So, if it’s going to be the centerpiece of your new ad campaign, or the intro theme to your podcast or whatever, and you don’t want other people using it, you have to hire a human musician. If the purpose of a piece of music is the music itself, you should hire a human musician because they’re going to do a better job than the machines will, at least with today’s models. That may change in the future, but for sure, today, you have to hire a human musician to make something that you know, as the opening sequence of your Super Bowl ad.

Christopher Penn 43:45
Yeah, you want a human musician, and if you want the human musician ecosystem to thrive, you have to put some money into it. That’s basically it.

John Wall 44:01
Yeah, no, that’s the right thing to say. Let’s support the artists and keep the interesting stuff flowing. But, by the same token, if you need a track to hide the background noise, this will get it done for you.

Christopher Penn 44:15
Exactly. Any final parting thoughts?

John Wall 44:17
John, yeah, we’ll be going on tour with our new album in three months. Once we get this all going. The seventies funk email survey tour is coming to a town near you. So, visit us over at seventiesfunkaimusic.com and we’ll see you there.

Christopher Penn 44:37
That is going to do it for this episode of “So What?”. Thanks for tuning in, and we will talk to you on the next one. Thanks for watching today. Be sure to subscribe to our show wherever you’re watching it. For more resources and to learn more, check out the Trust Insights podcast at TrustInsights.ai podcast and a weekly email newsletter at Trust Insights AI newsletter. Got questions about what you saw in today’s episode? Join our free Analytics for Marketers Slack Group at TrustInsights.org. Dot AI analytics for marketers. See you next time.

Need help with your marketing AI and analytics?

Leave a Reply Cancel reply

Subscribe to our Weekly Newsletter

Pin It on Pinterest