Suno (AI generated music service) – my experience so far (long)

3 points by squarefoot 1y ago ↗ HN

After some time testing the service, about two months ago I signed for a paid subscription on suno.com; here are my impressions. Need to split the post as it's too long.

0: A foreword - I'm not touching the whole copyright thing at all as I have zero interest in making famous artist X sing one of my songs, clone their music, or sending any musicians to flip hamburgers because I can't pay them to play for me; I write my own lyrics and, after having played for years in a rock band, just want to test the technology with an open mind. And.. yes, nothing beats the live/rehearsal room experience with real friends, instruments, sweat and lots of beer; no need to remind this old rocker about that:^)

1: Music genres available - There's lots of them, more than one can dream of, and they're so interesting to explore and combine; this aspect alone can have a huge potential in teaching music, if and when they'll implement a way to give the user more control over the AI. Unfortunately as it learns from the music available, it suffers from the "4 chords syndrome" plus if you're say a prog rock lover you'll have a hard time producing anything that is not some sort of metal in 4/4. I couldn't make it produce a 7/8 or 5/4, not even placing it along with "odd time signatures" in the prompt.

And consequently, prompting for metal and hard rock very often returns speed metal with annoying double speed drums. (to any suno engineers hopefully reading this: they're a rare exception, not the rule).

Apart from this and other possible quirks about other genres I'm not used to and can't comment, it's extremly rich and fun to explore. Also, it follows the indication if you write the era inside the prompt, that is, writing "rock, 70s, hammond pad, ..." will likely produce something that sounds like made classic rock of that era, and conversely makes possible things like "grunge, 70s, bagpipes" which of course will return something never heard before. That is what makes it interesting the most: sailing new seas!

Another problem: it's somehow obsessed with putting stupid synth solos in rock songs, and I don't mean a '70s type mono synth played Keith Emerson style which I would welcome in some tracks, but some obsessive repetitive loud patterns using multiple layered sawtooths like in many dance songs of a decade ago, that if you ask me, have no place even in dance music or techno (which I like). However I get them less and less with time, probably as a result of downvoting my own tracks that exhibit that problem; but that meant I smoked lots of precious credits for something that could be solved by having more control within the prompt or within lyrics. Trying [no synths] and the like of course didn't work. That brings us to the next point.

...

squarefoot 1y ago ↗

2: It's all fuzzy. Don't expect to ask for something in the prompt and hear it immediately; sometimes it happens but quite often it needs a few generations, so don't give up and be prepared to a nice surprise after multiple retries. This applies to everything, lyrics included; sometimes it sings the wrong parts, mixed up lines, words, or doesn't sing them at all, and sometimes it blends up words in strange new languages; that is also part of the fun, but can cost money.

2b: Hallucinations. They happen in the above form, or as bad sound quality (see below). If words are misspelled they can be corrected by the replace section function, however if as a result of hallucination say it sang a verse twice, deleting the unwanted one becomes hard because the AI now lost sync between music and lyrics, and when you attempt to correct that part it could operate at the wrong position. In such eventuality it's probably quicker and cheaper to extend the song from a point before the hallucination, so that everything onwards is overwritten, including the hallucination. Luckily however the function to replace a part costs credits only after confirming it, therefore one can retry indefinitely at no cost until the correct piece is generated then confirmed.

3: Sound quality - This varies from quite good to really nasty, and so far there's almost nothing that can be done to correct a bad sounding track or section : sometimes a song is heavily distorted (digital distortion, not like simple clipping), and quite often making a cover will generate another bad sounding track. Substituting sections doesn't always work. Another very annoying bug is the tendency to raise the volume of vocals and music with time, and that has nothing to do with different emphasis or changing anything in the way it sings, it just becomes louder and louder with time, usually after the 1st minute. If that happens (more often with some genres) again, here's nothing that can be done except making a cover and hoping. This is my #1 annoyance with Suno; pretty much 70% of my songs sound like crap because of that, although most of them start very clean, but then slowly become louder and louder; one can easily spot the problem just by looking at their wave plots.

4: User control - User can enter their lyrics or ask the AI to generate them itself based on the user prompt, and results vary from decent to sometimes surprisingly good. Also with original lyrics it's amazing how it finds the best way to fit them into a song, and I don't even write in English. Simply astounding! Control over song creation is quite limited however; one can give directions but chords insertion is limited to things like [G#m] before the affected line, and I think very basic rhythm figures can be entered in some way, but I still didn't test those features. A xoxoxo grid to enter rhythms would be a killer feature, but please, dear Suno engineers, don't assume everyone wants to make a 4/4 hip hop figure on a 16 step grid; my favorite drummer is Neil Peart and... yes, one can totally program YYZ-like things on a capable enough drum machine.

Personas are a very recent addition; once the user creates a song whose style (+instruments and vocals) they like, the personas function can be used to create a virtual artist that inherits those characteristics, so that following songs created using the same persona will be similar in style and sound. Be warned however that they seem to inherit low quality hallucinations as well. I've thrown away lots of credits wondering why a persona I created made great songs that I liked a lot, but all of them, and i mean really all of them, sounded like dog crap. If you get a bad song don't make a persona out of it as it probably won't be worth the effort, but keep it just in case they correct the AI so that it can be used to correct bad sounding songs and personas.

4b: Users songs, along with all pieces created to correct them, are thrown into a giant library that grows and grows, and ...

squarefoot 1y ago ↗

5: Uploading - Users can upload audio material, which naturally represents a source of panic among big publishers because they fear people can upload part of a hit (currently max length is 2 minutes) and use it to clone big name artists, which 1: can be done in other ways, and 2: is far from being the most fun way to play with the feature, and could be a source of legal problems although, I repeat myself, there are multiple ways to do that, both inside and outside of Suno. Here's how I used it: any of you recall of that C64 game called Commando with that really catchy soundtrack by that genius called Rob Hubbard? Well, I spent decades waiting for the day I could play it in real life, but never could with my old bands, so I grabbed the game audio, edited it to fit the most relevant parts into 2 minutes, and turned it into a disco-funky song with slap bass et al plus AI generated lyrics celebrating the old times playing 8 bit videogames. Yep, it works on chiptunes too, and it is damn fun.

6: User support - next to none. They reply to all mails and bug reports, but as you can easily guess it's all AI generated and contains lots of boilerplate text; however it seems there is something that takes notes about bugs, so it might be still worth reporting them. However I asked them a couple times about song moderation with a very specific question about potential trolling, and ther mail AI kept replying about writing PG-13 lyrics while the problem was completely different. Specifically, I asked if I publish a song with political tone, I mean, nothing inciting violence, offending or anything like that, but of course strongly biased, what would happen if a group of trolls would report the song as inappropriate just because they don't share the same ideas and want me silenced? They kept replying about a moderation team, an I replied back that in case the team is an algorithm, then it could be easily gamed, like a bunch of flat earthers could silence a scientist just because of mere numbers. Still didn't receive any meaningful reply on the subject, so I'll keep my profile private for now, possibly sharing my song elsewhere. Better be sure before risking to lose everything Google/Youtube-style.

I'm sure I forgot something I'll recall later; in that eventuality I'll add that on this post. What are your experiences with Suno or other AI music generation services? So far it seems Udio offers much better sound quality, however it's still limited to 2 minutes (I'm aware of concatenating pieces, but it's not ideal, not even on Suno), but they progress at such speed that everything could change from day to day.