Sold my Soul to the Devil

[richiebee]

I fed Suno one of my songs the other day. I'll leave aside the loss of rights to my own song. That's a different topic.

I asked it to do some covers for me, and what it came back with shocked me. People say it has no soul, just banal mediocrity. What it did to my cover though was absolutely brought it to life. It kept things like the notes of the guitar solo in tact, but made it sound like it was actually played by a guitarist who was having a good time. It tastefully altered some of my chord voicings in a way that I might expect from an orchestrator with a much better understanding of theory, and a command of the task than myself and gave it a great overhaul. The audio quality left a lot to be desired, but ultimately, it gave me a whole bunch of ideas to use to improve the arrangement of my song for a re-production.

I don't see any point in getting it to write a song for me - I have the radio if I want to listen to someone elses song, but for brainstorming my arrangement, it did a great job.

 

[David Cuny]

What it did to my cover though was absolutely brought it to life. It kept things like the notes of the guitar solo intact, but made it sound like it was actually played by a guitarist who was having a good time.

Yeah, it really shows the potential for what neural networks can do for virtual instruments.

It's a shame Suno got there by treating music on the internet as free training data.

 

[Nekujak]

I have a young friend who's a filmmaker (commercials, corporate videos, short form videos). Since I did a lot of video work once upon a time, he and I often discuss film, cinematography, video production, etc.

The other day we were talking about lyric videos, and afterwards, he sent me a lyric video he recently made for a singer-songwriter friend. The video was beautiful, but the music and the singer's voice (female) was absolutely stunning.

I played it for my wife, who's a professional singer, and we were both impressed by the singer's voice and how well-produced and polished the track sounds. My wife even said, "She must've spent a lot of money on that production."

You know where this is headed...

About a week later, I saw my friend and mentioned how impressed we were with the video, and that we were blown away by the music. He looked down and sheepishly muttered, "That's not her voice or music, it's AI." It turns out she's not a singer - she's not even a musician. She's a schoolteacher. She wrote the song - chords, lyrics, and melody - and fed it to Suno.

It was a bit of a gut-punch on couple of levels. I didn't like that my wife and I were led to believe this was a real human performance. I know my friend wasn't deliberately trying to fool us - he was simply sharing a video he made, and didn't mention the music because it's not his area of interest. It made me realize we're probably already inundated with AI-generated music in our daily lives more than we're aware of.

But what I really didn't like, is that my wife and I were so impressed by the track. It just made me think, "Why bother?". Why bother learning to use a DAW? Why bother carefully crafting an instrumental arrangement? Why bother struggling to record a decent-sounding vocal? Why bother spending the time and effort to record and mix a track?

It's one thing to type a prompt and have AI spit out a dozen songs for you. There's clearly no musical creativity involved.

But it's a slightly different matter when you actually put in the work to write a song, and then hand over the production tasks to AI. As much as I despise generative AI, this scenario is now starting to sound a little less slimey to me.

You could argue I'm taking work away from musicians, producers, and engineers. But am I really? If the alternative is me producing the track in my DAW, that's also technically taking work away from musicians, producers, and engineers. The net result is exactly the same. The only real difference is I'm either just the songwriter, or I'm the songwriter, performer, and producer.

And yes, you could split hairs and say that by using AI to produce the track, I'm robbing sample library and plugin developers of revenue. But again, am I really? I spent many, many years and thousands of dollars buying sample libraries and plugins. They already got my money, and I already have all the production tools I'm ever going to need - I rarely spend any money on gear anymore. So again, same net result.

And you could argue that AI is robbing artists of revenue by training on their music without permission. Well, that was definitely true at one time, but as of last fall, Suno is now owned by Warner and only trains on the Warner catalog. And the same is true with Udio and UMG. So again, same net result. In fact, by using my DAW, I'm technically denying money to Warner and/or UMG artists who could potentially earn something when their music is used in a generated track.

My brain hurts thinking about all this. But it is causing me to re-evaluate my stance on AI's role in the music production continuum.

 

[richiebee]

It's a shame Suno got there by treating music on the internet as free training data.

Fair point if they did indeed steal people's music. That is if they used illegal methods like evading a subscription (whether literal or by bypassing ads that pay revenue for music licensing) to a music streaming service. I assume this is the case since the whole industry is up in arms and no one seems to defend Suno in this regard, not even themselves. But if they did pay for a subscription, or used music that was already "free" to listen to, then they just did what you or I could do to analyse people's songs. They just did it much quicker.

 

[David Cuny]

My brain hurts thinking about all this. But it is causing me to re-evaluate my stance on AI's role in the music production continuum.

I'm not interested in generating music with Suno. But using it to create vocals and transform instrumentals is really tempting.

When I put together Down That Long Road, I used SynthV for the vocals.

But it would have made a lot more sense to use Suno for the final vocals. It would have been a voice better suited for the genre, as well as being more expressive and realistic.

The main issue was that I didn't have a subscription to Suno because of issues I've already alluded to, although licensing agreement with Warner/UMG addresses that to some degree.

I'm also not thrilled that since it's not trained on isolated instruments, extracting vocals and instruments requires stem splitting and all the inherent issues with that.

But even limiting myself to that, I'm still constantly reminded that Suno's primary function is to generate music for me. There's an option where I can give it samples of my voice, so it can create a voice model so I don't have to sing. And there's a function to upload music I've created, so Suno can automatically generate music, so I don't have to write music, either.

So I'm still very conflicted about this.

 

[David Cuny]

Fair point if they did indeed steal people's music. That is if they used illegal methods like evading a subscription (whether literal or by bypassing ads that pay revenue for music licensing) to a music streaming service.

They simply treated any music they could access as fair game for use as training data.

But if they did pay for a subscription, or used music that was already "free" to listen to, then they just did what you or I could do to analyse people's songs. They just did it much quicker.

That's a false analogy. As Adam Neeley noted, it's like saying that a submarine is just like a person because they both swim underwater.

When the music wasn't posted "free" for corporations to use as training data. It was posted so that people could listen to it. Corporations took advantage of the lag between law and technology.

This "they're learning like people" is a lie promoted by AI companies trying to claim that laws intended to regulate the behavior of individuals should apply to what companies are doing on an industrial scale.

AI doesn't hear like humans, or remember like humans; it doesn't learn like humans or even create like humans.

There is no human who can analyze millions of audio files, remember them all, and then create songs that mimic the instruments and voices in those files by writing zeros and ones to binary files.