Hacker News Reader

Comment by adriand | Hacker News Reader

This seems like a good decision, although, is there a good way to tell if music is AI-generated? I assume that some of the music that's showing up in my Spotify feed is AI-generated but I've never noticed.

View on HN

So there's really accurate ways to detect "pure AI". The AI music detectors out there are mainly looking out for production things:

-a flatness to the EQ spectrum that you wouldn't get out a properly mixed and produced piece of audio

-no good stem separation, so no per-source eq (relates to above point)

-change BPM mid-song

-unnatural warbles at the end of every phrase

-vocals will have these weird croaky voice cracks, or sound scratchier and raspier

There definitely are tell-tale signs of "pure AI" in audio, but it becomes a lot more nuanced when any sort of secondary mixing/mastering/compression happens (which is the case 90% is the time in the real world- anything on YouTube/Spotify get's compressed).

> I assume that some of the music that's showing up in my Spotify feed is AI-generated but I've never noticed.

A lot of it is now, and it's frustrating to me. The worst part is that I'm not actually anti-AI-music. There's one or two "groups" ("producers"?) I've found where it's clearly AI but they've put a lot of work into making something worth listening to, but Spotify seems to have a "this sucker will listen to the cheap stuff" flag and now I'm drowning in tracks from people who paid for Suno and think that's enough.

I'm not sure about the flag :D but I've had pretty good results with always making sure to flag every artist in AI suggestions with "don't play". You have to visit the artist profile pages to do it though, so it remains a cat and mouse game, but I think that as long as they don't prevent it by force, doing this tends to improve the AI suggestions (of non-AI music).

Similar to YouTube slop.

If that would stop working, I'd cancel Spotify again.

Speaking of YouTube slop, I think Spotify has had its own system of preferring cheap muzak from labels they support since before GenAI music even took off, I think. Example label: Firefly entertainment (IIRC)

At least for the current AI music generators, it's pretty easy to tell by ear that it's AI generated. Everything is just a little off, especially the higher frequencies. Vocals often sound indistinct, like an unholy amalgamation of thousands of people are singing instead of a single person.

I think it would be very difficult for most people to tell that songs are generated by Suno 5. There are some interesting anomalies I can see in the spectrum and mid/side channels, like Suno music often has very little information in the side channel (what happens when you subtract the left and right channels from each other). You also commonly see the eq curve of the rhythm section shift over time throughout the song - like drums will sound normal at the beginning but end up sounding kind of under water by the end, but they are quickly improving these things. But to the layperson, many of these things are completely invisible. The most obvious tell, IMO, is the cadence of the lyrics.

Does this apply to all genres or just highly produced popular music? I would not be surprised if I failed to detect an AI song as background in a television commercial, but it is difficult to imagine that anyone could fail to pick out an AI impersonation were you to slip one in to a record like 'João Voz e Violão.'

It really depends on the style, yes. You could probably slip one into any modern pop/dance/club/EDM album and no one would know as long as the vocals sounds like the performer. For styles which are very unique with that sort of imperfect human touch that makes music so enjoyable, it would likely be obvious, at least at the moment.

It's getting very hard. At this point, lyrics are the biggest giveaway. AI generated lyrics are always awful and the delivery feels very stilted.

Currently it sounds like it's been through an allpass/comb filter. Complex parts, while spectrally there, do not make much sense as a real sound. Probably audio analog of the "finger salad" of early image models. I do not count of being able to tell one from another in a few months.

It sounds like a moral stance on its face, but honestly they probably wouldn't care if someone posted a reasonable amount of AI-generated music that was high quality enough to gain a following of listeners.

This is likely a stance to prevent an individual from producing thousands of AI generated tracks and attempting to flood the zone for anyone browsing and searching.

There's a lot of music on Spotify for example that tries to latch on to current trends in an attempt to get pulled into search results and recommendations.

One day soon many musicians will be using AI assistance, and many won't tell you for fear of judgment.

It's like that with code and art.

Purely AI anything is garbage. But AI tools in the hands of people who know what they're doing are just faster scaffolding and better plywood to build with. The framing is still mostly human expert.

> One day soon many musicians will be using AI assistance, and many won't tell you for fear of judgment.

Word on the street here in Nashville is that it's already the case. The songs getting published aren't AI-made, but there's AI assistance.

Auto tune uses a form of "AI", and has been used by most pop singers for a decade.

Auto tune is not what makes a song a hit. What make a song a hit is the fact that the person who injected the use of auto tune has taste.

This seems to fly over the heads of many. art is about taste.

The same argument applies to AI generated or assisted music. Anyone can write a prompt and get a song. It takes judgement and taste to pick a good song and choose to publish it.

"New technologies have the tendency to replace skills with judgement – it's not what you can do that counts, but what you choose to do, and this invites everyone to start crossing boundaries." - Brian Eno

AI seems to struggle to produce counterpoint, especially in vocals.

Reply on news.ycombinator.com