🎵｜stable-audio

[south dawn]

Introducing Stable Audio: Stability.AI’s first AI product for music and sound generation! 🎧

“Ready for something new? Welcome in Stable Audio!”

Stable Audio is a first-of-its-kind product that uses the latest generative AI techniques to deliver faster, higher-quality music via an easy-to-use web interface. Stability AI offers both a basic free version of Stable Audio, which can be used to generate and download tracks of up to 20 seconds, as well as ‘Pro’ subscription, which delivers 90-second tracks that are downloadable for commercial projects.

Stable Audio is the first music generation product enabling the creation of high-quality, 44.1 kHz music for commercial use via latent diffusion. The latent diffusion architecture uses audio conditioned on text metadata as well as audio file duration and start time, allowing for control over the content and length of the generated audio.

We welcome you to give Stable Audio a whirl for yourself, and provide us your feedback! We want to hear from YOU!

GET STARTED:

http://stableaudio.com/
#🎵｜stable-audio: For discussion, feedback, prompt-sharing, and outputs surrounding this release!

[prisma reef]

in addition to music it can also generate sound effects! The user guide has examples of both music and sound effects generated if you scroll down https://stableaudio.com/user-guide

[wind hedge]

Best audio generation so far. While I appreciate jukebox's willingness to get a bit chaotic and unpredictable, just spent the last hour generating useable stems with stable audio

My god this is amazing

[daring hamlet]

We're thrilled you like it 🙂

Please feel free to share your outputs and any great prompts you discover - we’re really curious to see what works well. We've only just started to scratch the surface of how best to prompt the model.

[cloud geode]

I just gave it a whirl! while I'm happy with the improvements in Audio quality the composition of the pieces is... well pretty bland usually.

I've written music for decades though so maybe my standards are higher than most !

As a musician the holy Grail for me would be being able to influence the generation by providing a .midi file or a reference audio file to guide the structure of the piece.. sort of like how controlnet works in stable diffusion? I'm sure that's not applicable in terms of coding but you get the idea.

It's really difficult to thoroughly describe the structure of a piece of music (impossible?) Through natural language so I imagine that will be a challenge as well. All In all pretty cool !

[shut saddle]

Fauno is going to be hosting an office hours this week discussing the project so if that sounds interesting, join the Harmonai discord server so you can listen in.

[cloud geode]

I'd love to if I have the time I'll keep an eye out

[compact condor]

I have a question regarding copyrights and selling AI generated music. I read https://www.stableaudio.com/terms and just to be clear, is it legal to sell generated music to services like SPotify, Apple Music etc..,?
Of course Professional account is the first step.

[prisma reef]

what you're describing would be a "commercial music release" so per this chart

that gets a checkmark for a "Professional" tier account

(I'm not a lawyer though, if you're concerned about the precise word of the law, especially if money's involved, contact a professional lawyer to be safe!)

(also bear in mind you are bound to the terms of the relevant platforms you're publishing to, not just stableaudio's terms)

(also you probably don't want to just generate on stableaudio and then go sell it. It's a tool for musicians not a one-stop-shop!)

(also copyright law around AI in general is a slightly weird topic so doubly contact a lawyer to be safe before going through with plans to sell stuff)

[compact condor]

Ok, thank you for your answer 🙂

[warped vine]

That might be enough to dissuade 90% of us from trying to make money, haha. Lawyers aren't cheap!

[compact condor]

I took all the answers as No 😂

[prisma reef]

I mean the answer is "yes" and literally says so on the homepage in the screenshot there but i'm not a lawyer and am too scared to just say "yes" if the topic is anywhere close to legal issues

[small harbor]

Trying to make some Noisia style d'n'b. Not easy. First prompt:" Noisia style drum and bass, Heavy and Distorted, Complex Drum Patterns, Dense Soundscapes, Dark and Atmospheric, Technical Mastery" As can be heard, the quality is garbage.

Second try with simpler prompt: "Noisia style drum and bass". Again very lo-fi sound but the idea is there, so let's continue.

[charred oak]

Here is my first try with a kind of meditative music ..

[small harbor]

Next prompt: "Neurofunk drums, 174bpm, crisp sound, professionally mastered, high quality" Now there is more high frequencies, but the drums sound very bad. Yet the idea is ok. So let's try something else.

As the previous prompt had high frequencies, but it was just soloed drums, let's add some bass: "Neurofunk drums, sub bass, growl bass, 174bpm, crisp sound, professionally mastered, high quality". Ok that is just silly, i'll give up. Conclusion: Soundquality is not much better than in Meta's Audiocraft tha i run locally. Note that i have used a ton of Harmonais Dance Diffusion to train my own audio sample models and used all other AI audio generators a lot.

This promp is a bit demanding, but still...

[sly fjord]

just gone for pro but my profile isn't updating and generations stopped working... also it won't play my last generation... any help?

[small harbor]

The site is buggy, refresh your browser and you see the generations

[warped vine]

I saw a prompt I liked for one of the samples on the model card. But my first two generations aren't anywhere near as good. The prompt was: epic trailer music intense tribal music and brass

[sly fjord]

tried that, no dice... I guess i'll wait

some spacey sound fx

[unborn gate]

Is site up and running?? I am getting loader on generate page.. I can Input prompt.

[sly fjord]

not running for me

must be getting a lot of requests

[oak umbra]

马斯克

[unborn gate]

Yeah its loading on generate page 😢

[sly fjord]

same

[unborn gate]

Well that's sad.. Let's hope it up soon..

[small harbor]

It took like 5 min to re-load the page. Just wait when the AI youtubers make a video about this, Then the site is gonna really crash.

[unborn gate]

I am one of them 😂 I was making video and it crashed on my first prompt and now I am stuck 😂

[small harbor]

Yeah, the AI youtubers have problems if the site does not work.

[bleak locust]

Hey. There is a lot of traffic on the site currently. We are looking into fixing this asap.

[unborn gate]

It's problem for everyone but I understand what you are trying to say..i would just take a step back

[sly fjord]

so far, i like it for sound effects, but music has been disappointing in first few tries (admittedly i did not go easy but when i tried something "techno".... it was a hard pass)

what I don't understand though... is Stable Audio different from what Harmonai is doing with Dance Diffusion?

[small harbor]

Sound FX also have bad artifacts on the sounds. Harmonai is also Stability AI company like Stable Audio. Probably same people.

[sly fjord]

the SFX were cleaner than the "music" i was getting 🙂

I have been using Suno for a couple days now... it's better, altho also imperfect of course

(for music)

[small harbor]

It depends on the prompt what it generates. Music is from crappy sound library that is not very professional sounding.

[sly fjord]

I asked what they use under the hood but got no reply

[small harbor]

They use Dance Diffusion code

[sly fjord]

aha!

it sounds a lot different from Stable Audio

different artifacts

and way more coherent compositions

but again, i only tried like 10 prompts so far

and 200 in Suno 🙂

[small harbor]

They used DD to train the model. I'm waiting them to release the updated DD code as it's really old. DD can generate high quality sounds without any artifacts. But it's for single shots. As soon as it's loop or longer sound, the quality goes down.

[sly fjord]

I played around with the Colab before, been waiting for a nice interface

[small harbor]

I run DD locally with 2 different GUIs.

[sly fjord]

interesting

don't have the gpu for it personally

cloud computing's so convenient

[small harbor]

I have now tested Stable Audio against META Audiocraft and so far Audiocraft is winning. AS has it's moments, but it's hard to get usable results. With Audiocraft Plus i can daisychain prompts up to ten prompts and generate 5 minutes of audio.

[sly fjord]

oo I haven't checked the Meta one

is it public?

[small harbor]

Yes, there is public demos and collabs for it. I run it locally

[sly fjord]

shall investigate thanks 🙂

[small harbor]

Here is a prompting guide for SA: http://www.jordipons.me/on-prompting-stable-audio/

This explains a lot: "Because part of the dataset contains MIDI instrumentation, some generations can sound like MIDI instruments."

[sly fjord]

seen yeah thx 🙂 👍

[sly fjord]

that blows a bit

https://huggingface.co/spaces/facebook/MusicGen tried this before, wasn't so great iirc

[hollow dune]

how to use stableaudio?

[sly fjord]

website's being fixed... and Drommer-Kille just posted a guide

[hollow dune]

can we run it locally

[sly fjord]

I take it back, it's not that bad, some clear moments here and there

[hollow dune]

i know fb's musicgen, does staleaudio have local ver

[prisma reef]

https://stableaudio.com/faqs

[cloud coral]

someone else can't generate?
site is not loading, cleared cache already

[modern charm]

give it a time, probably getting swarmed right now and they are working on scaling it further

[charred tendon]

I thought maybe if I subbed to Pro it would maybe allow access to Generate as like a fast track to use the service....nope. lol

[scenic marten]

I had the same thought haha

[shut saddle]

This was quite a surprise to wake up to!

[scenic marten]

Anyone tried to generate melbourne bounce music? Any good hits?

[celest magnet]

Is it gonna be opensource?

[modern charm]

https://www.reddit.com/r/StableDiffusion/comments/16hkivg/comment/k0echmw/?utm_source=reddit&utm_medium=web2x&context=3

We are building the open dataset version now that we will release along with the trainer etc per the research blog.

Also continuous improvements to the platform and features so you can fine tune your own models, create music videos and much more (done when they are done).

[charred tendon]

When they fix their website...I will let you know. lol

[celest magnet]

Is the site broken it won't load the generate page

[charred tendon]

Supposedly they are working on it, from what I read further up in this channel

[celest magnet]

It sounded pretty great from the demo

[hybrid star]

AI Music hmm

[charred tendon]

Yeah, I agree. That's why I wanted to play around with it for a bit see what I can create. Hopefully the generate part of the site goes online again soon

[iron owl]

And what about this then in the open source version?

[scenic marten]

good stuff, send me some bangers

[charred tendon]

Ok update. After maybe 20 mins I think of loading the page, I saw what the Generate page looks like. Just wanted to test it to see what would happen if I reloaded the page...well, its loading again. lol So yeah, its probably just bogged down or being fixed at the moment.

[unborn gate]

Happening for last 2-3 hours.

[eternal remnant]

Guess I'll wait for a bit

🙂

[charred tendon]

Oh. Well hopefully they get it ironed out. I can't even get the site to view me as a Pro member, even when I cleared the cache and logged back in with my email account that I used to pay for the sub. So yeah, they clearly have a few things to iron out till the site is working as it should. It shouldn't be much longer I feel

[river latch]

I'll second this.

[shut saddle]

Pretty safe to say that you'll be lucky to get a gen or two today.

[odd geyser]

got kicked while it was generating 😭

[unborn gate]

Hopefully this get resolved soon..

[odd geyser]

D1 Error

[shut saddle]

Yeah I'm guessing they are overloaded.

[unborn gate]

getting diffrent error now.

[unborn gate]

It was expected and it's gonna increase in next few hours more..

[shut saddle]

Yeah I was hoping to try it out because I'm gonna get a lot of questions in the main Harmonai discord.

[harsh hatch]

How long does it take to generate one audio?

[odd geyser]

I just want Haydn trumpet concerto 4th movement

[shut saddle]

So far the answer is D1 error.

[odd geyser]

guys can you all stop generating for 5 mins so I can get a 45sec audio tysm

wtf it uses up the attempts even when it fails?

[shut saddle]

I got a "cant generate more than 1 track at a time" so I'm guessing yours will pop up once the server finally gets to yours.

[odd geyser]

I got another error

when I press download this happens

I'm on mobile, is that the issue?

[shut saddle]

🤷‍♂️

[torn zinc]

Is the model downloadable?

[charred tendon]

To my knowledge, it's not open source

[torn zinc]

Thanks!

[shut saddle]

Yeah they haven't released a model yet.

[merry raft]

the duration limit for pro is way too short to be viable imho

instead of 90 seconds, it should be 240

also choosing mood and instrument is nice, but what about bar and chords structure ?

[crimson turtle]

Looking forward to it working:
Never got my first generation but a credit was deducted.
Waited an hour, submitted a new prompt but it said I can only have one prompt at a time -- i.e. the prompt that hung up is blocking further access.
Just got this message now, so contacting support:

[stark pawn]

Just got this error as well

[foggy elbow]

they should throw in an extra 100 credits for us who bought it and cant use it

[shut saddle]

Same 🥲

[fresh falcon]

hi

[crimson turtle]

@foggy elbow based on how DreamStudio was operated, I suspect they will be happy to be generous in resolving issues of lost credits.

[wind hedge]

Would upgrading to Pro circumvent the traffic issues?

[scenic marten]

no

[wind hedge]

😭

[scenic marten]

I recently upgraded too and it didn't help at all

[wind hedge]

That's a deal breaker for me then

I literally saw the ping at like 4 am and then used it immediately, so that's why I was able to generate most likely

[scenic marten]

try again tomorrow 😭

[prisma reef]

I don't think the team anticipated quite how much attention this launch would get lol

[fresh falcon]

should’ve released it quietly

[shut saddle]

It's a stablity based audio project. I figured it was gonna get some attention.

I think there's going to be a few people thinking this means there's an open source model available and that's not the case.

[iron owl]

Speaking of, any rough time frame on becoming open source? Because they did say it will be available

[shut saddle]

Soon is the best answer I can give. I don't really speak for Harmonai or the team. I'm just a discord janny.

[tame canyon]

I think SAI is underestimating just how popular music generation can be, I wager it will surpass txt2img in the future

[iron owl]

Huh. I remember setting up their dance diffusion stuff, was neat but far too resource intensive for anything local, and the architecture didn’t really allow for full songs to be generated, only snippets, and it was more like random loras as opposed to a general text to audio model

[shut saddle]

Fauno has made a lot of progress since then but hasn't released the code to the public.

The oficially released stuff is close to a year old at this point.

[iron owl]

Quite unfortunate, I honestly believe that the more locked down some ml project is, the less innovation will be in the field it is in

[tight anchor]

We're working hard on the open source code and models 🙂

The library we'll be releasing is the same library we used to train Stable Audio, and will also support our newer versions of Dance Diffusion

[odd geyser]

site is completely down for me 😦

[tight anchor]

Yeah, it got hugged to death with all the traffic, we're working on getting it back up and running smoothly

[odd geyser]

[shut saddle]

It means we love it already 🙂

[tight anchor]

I can't wait for y'all to get in there and try stuff out!

[merry raft]

you can always try the meta alternative, it s prett much the same

[tight anchor]

We've also got thumbs up/thumbs down buttons on the site, which we'll be using to improve the model, so if something sounds good, smash that thumbs up button!

And if it's bad, smash that thumbs down button

[sick cipher]

@pliant chasm @elfin jackal @neon silo
Someone from stability DM me. Want to share a security issue.

[tight anchor]

You can message [e-mail] if you see issues

[charred tendon]

Thanks for posting this. At least it gives some sort of official update.

[shut saddle]

If we have queued an audio clip will it eventually generate or will it timeout?

The one I queued seems to be persisting between refreshes so I'm assuming it will eventually generate.

[charred tendon]

On the site, I still have a 45 second clip that's still in "generate mode", so yeah, not sure

And its been at least an hour

[shut saddle]

Same.

[odd geyser]

the meta one that I just tried was pretty bad tbh, hopefully this one is better

[shut saddle]

Personally, I'm banking on the assumption that an open source model will be so much better once it releases.

[trail hazel]

Beautiful 🙄

Gonna have to learn to prompt now lol

[shut saddle]

The model is probably not going to recognize copyrighted music. Just saying.

[wind hedge]

So I can't generate things in the style of an artist?

[shut saddle]

Not likely. There's pretty strict restrictions on music and how it can be used.

[trail hazel]

I guess asking everything at once I shouldn't be surprised it sounds like it's getting to play 3 songs at the same time 😅

[trail hazel]

"Deep dirty rock with electric guitars, drums, electric bass, a string quartet, a horn quartet, synthesizers, and a high pitch vocalist. Rhythmic. 7/8. 140bpm. Interesting harmony."

Also, seems like it only knows 4/4.

[spring girder]

it's actually running on some poor 3070 in someone's basement so the outages make it look more popular

[scenic marten]

I have a 3090 for lease if anyone wants it

[charred tendon]

Please tell me this is not real. lol Yes, Im gullible sometimes. lol

[prisma reef]

that's not real

[proud dragon]

Hello class

[shut saddle]

My audio track is stuck and isn't generating. Has anyone had success or is my account bugged now?

[river latch]

Yeah, mine has been spinning for the last 30 minutes.

Though I'm sure that's just what happens when someone like me ignores the warning label telling us to come back tomorrow, lol.

[shut saddle]

I didn't have the warning label at the time heh.

[inland wasp]

How to report a bug? My generation is stuck since half an hour ago and I can't submit new prompts ☹️

[shut saddle]

You aren't the only one. They haven't mentioned a fix yet.

[tacit jackal]

i hope there's a way to regain the credits we lost on requests that timed out. congrax everyone at Stable, this seems amazing 🙂

[shut saddle]

Those credits are 20 per day so even if you miss out don't worry.

[tacit jackal]

says here it's 20 per month: https://stableaudio.com/pricing-with-account

[shut saddle]

Oh my mistake. Month.

[tacit jackal]

still pretty generous for free plan

[shut saddle]

Yeah hopefully you'll get reimbursed.

[tropic cape]

Is there a Stable Audio API?

[daring hamlet]

Not yet, but we're working on it

[tropic cape]

ETA?

[river latch]

[reef mica]

What happened to "open-source"????

[shut saddle]

They haven't released the audio model yet. Relax. Let them test it out for now.

[modern imp]

I gave a prompt for a 7 beat per measure bass, but it gave me 4/4. It sounded pretty cool, but not what I had wanted.

[stark pawn]

For something called "Stable Audio" the website certainly is not stable at all lmao

Not bashing it though, I'm really looking forward to using it 👀

[copper glen]

Subbed, but doesn't seem to be reflected on my account. How long should it take?

[copper glen]

Already relogged and refreshed and whatever, still says "Upgrade to Pro" and clicking that says I already have a sub

[plain knoll]

"hardcore" which is a genre of music gets removed from prompts :/

[winter halo]

I'm documenting the limits of Stable Audio.
I downloaded one so far, I'm waiting for others as it is overloaded.
Already it seems it is far, far, very far behind MusicLM by Google Research. Note MusicLM fell apart finally as 3 months hit and now repeats itself, thankfully I got in all but my sadly hardest test early enough.

You can see those early-only tests, and how far ahead MusicLM is by listening to as many of my top picks below as you can, or at least the rin kagamine ones, lost lava world ones, and large diamonds in advanced starship ones.
https://www.reddit.com/r/singularity/comments/13h0zyy/i_really_crank_out_music_tracks_with_musiclm_this/

Also, just below is my first test with Stable Audio so far:

@south dawn

MusicLM released in 2023 January also. I'm just showing what exactly yous are up against.

[wary sky]

in my opinion suno chirp is still better in genral. but stable- audio has a better stucture becasue its trained on coplete songs and it can make longer songs. this is a example from chirp. and yes chirp can also do lyrics. but the fundamental sturcure of both ais are different. the chirp model works simular to a laguage (gpt) model and the stable audio more works like stable diffusion.(diffusion)

[delicate stirrup]

hello there! is there a way to install stable audio locally?

[tidal lava]

That's it!!! Rename stable audio to Unstable Audio

[charred tendon]

Looking up in the channel, you will see that they mentioned it's not open source with what is available right now

[shut saddle]

This is web only at the moment. They haven't released any weights yet.

[copper glen]

beautiful

[heavy arch]

If we ever get a selfhosted open source one. Feels weird to sell subscription's to a test, i hope i am just paranoid after openAI going full corporate for the cash grab

[wind hedge]

damn, even 2 am and can't access the site

[shut saddle]

Fauno intends on releasing one. I don't see this changing anytime soon.

[winter halo]

Outputs are choppy (I think, I might be wrong).
Does Pro Plan make audio better????

HOW BETTER 🙂

[plain cloak]

Trance, Ibiza, Beach, Sun, 4 AM, Progressive, Synthesizer, 909, Dramatic Chords

Trance, Ibiza, Beach, Sun, 4 AM, Progressive, Synthesizer, 909, Dramatic Chords

[winter halo]

are the WAV files it allows you any better? Right now outputs sound squeaky! and choppy!

[winter halo]

i got some tests in just an hour ago, got 4 out of it, there's a way to do it xD

[copper glen]

Site seems to be working right now

banger

[sly fjord]

not working for me still but at least my credits have been added

[tidal lava]

What was the training corpus for Stable Audio?

[scenic marten]

I haven't even been able to generate a single song for over 24 hours

[copper glen]

It's generating immediately for me now

i'm so sorry 🥲

[sly fjord]

ok now it generates, but the sound player doesn't work... at least the download works so yeah, getting there

[copper glen]

Ok now it's broken for me

We takin turns?

[copper glen]

spooky

[sly fjord]

the Monolith is near, i can sense it

[tidal lava]

Will it be possible to finetune the model with my own instrumentals and vocals once it's open sourced?

[tight anchor]

Should be possible, yes

[knotty quiver]

It's a real mess. It's like 10 tracks playing at the same time.

[deep thistle]

after 5 attempts and a long wait... 2 were successful, the sound and the song in general are very bad, clearly this model will need more work!

[urban venture]

Can the audios be rebroadcast with editing?

[elfin zenith]

how long does it useually take to generate?

[elfin zenith]

wth XD

[cedar narwhal]

As a 29-year jazz arranger and music teacher, I cannot WAIT to make tons of new music. Looks overloaded right now, testing with Bach inventions to see how it gets counterpoint, and will then test Mozart and Debussy to see how close to their music it can get. My whole education was studying advanced music theory and getting kids scholarships to college, this looks beyond awesome and right up my alley for being able to use at top level.

[placid pewter]

This sounds like a Phish show I once went to.

[placid pewter]

that's not bad!

[cedar narwhal]

simplify your prompt in some ways, but be specific in others - AABA form, less instruments, etc

[deep thistle]

far away from breakcore, it`s so strange

[unkempt steeple]

I am trying to see how good is with latin music, one of my first tries with salsa:

[little bone]

When will it be open source

[shut saddle]

There's going to be an office hours today you could ask that there.

[compact condor]

Audio prompt was: drum and bass, fast, energetic, melodic, 160 bpm

[exotic raven]

My first try/creation, that's so nice, I love it.
Relaxing Irish flutes, love melody.

[tight anchor]

Triumphant orchestral brass fanfare for the return of the glorious king

lo-fi hip hop beats to relax to

[next gyro]

ooo this one is a jam

[tight anchor]

3/4! Nice!

Occasionally the model moves away from its 4/4 bias

[tidal lava]

This sounds great af

[tight anchor]

lo-fi bluegrass to till the fields to

[next gyro]

tavern fantasy music, lute, flute, fantastical

[tidal lava]

How long does it take for you to generate?

[copper glen]

A minute or so

Sometimes needs a page refresh

[sharp viper]

Thx haha 🙂

[tidal lava]

Thx, that's what I needed lol, a refresh

Vivaldi, Bach, baroque violins and cellos, combined with Michael Jackson singing, with Jimi Hendrix solos, trap drums

[tidal lava]

I mean, I didn't expect it to be a good combination of genres, but this sounds terrible lol

[shut saddle]

Harmonai office hours is starting if you guys wanna join us.

[copper glen]

Where's that?

[shut saddle]

It's the Harmonai discord server.

[mighty dome]

I read the rules before posting this to see if there was anything about putting this link here, and I didn't see anything, but please delete it if it's against the rules.

Stable Audio seems to be working a little better. I'm streaming stableaudio generations. I'm not monitoring the chat. The prompts are in the file names, and were themselves generated by Claude 2 --

https://www.twitch.tv/endalarius

[plain cloak]

Guitar, Russian romance

[crimson turtle]

https://tenor.com/view/muahaha-evil-laugh-evil-laugh-futurama-gif-4133163

IT'S WORKING!!!!

[exotic raven]

a bit squeaky but still very nice

[crimson turtle]

This track could benefit from further development by @covert patrol

[covert patrol]

oooOoOoOo

[modern charm]

https://www.reddit.com/r/StableAudio/comments/16imtrs/catchy_kpop_hook_postrock_guitars_drum_kit_bass/

[covert patrol]

finally got it worken eh

[copper glen]

pro sub still not processed on the site 😦

[tender otter]

any news on when the audio model is released?

[wind hedge]

I wonder how much vram it would need

[tight anchor]

We're still working on the open-source models, hoping to get something out in the next month or so

Basically "when it sounds good"

[tender otter]

based on the demos I've been listening to, they are really good

will the dateset be released?

[wind hedge]

Just checked and I got 20 free generations back I guess lol

[prisma reef]

https://stableaudio.com/faqs

[tender otter]

yeah but will the set be released (like LAION for SD)?

[shut saddle]

AudioSparx owns the initial audio dataset for the stable audio model, so that's not what the open source one is going to be trained on as far as I remember.

[tight anchor]

Correct, the AudioSparx model will not be made public, we’re training open-source models on different datasets.

[neat python]

soooo ive been waiting 4-5 horus for mine to load

free version

abnormal, yeah?

[tight anchor]

Have you refreshed?

[neat python]

no im scared too lol

should i?

[tight anchor]

Yes

[neat python]

looks like just a loading screen now

oh!

i see there is a play button n such now. but taking a bit to start playing? on the site. maybe il try downloading

ohp. looks like takes a while too

this sounds horrible

yikes!

[tight anchor]

It's certainly hit and miss, make sure to hit the thumbs down on the ones that are just awful

[neat python]

will do

if curious

[keen hare]

I'got a "Sorry, you have been blocked
You are unable to access stableaudio.com" message when I first try to enter the website

[wind hedge]

Lol generated clips started disappearing, some doubling, and now I have another 20 credits again lol

Anyways

This put a massive smile on my face

That face when the AI drummer has a cymbal stack

[civic rampart]

my beat is crazy no cat mewing tho :(

[river latch]

Ok, call me impressed... And I'm hard to impress when it comes to AI audio generation. SF gets the most important thing right when it comes to generating audio, and that's following your prompt--quality and cohesion comes after that. I look forward to what y'all can do going forward. Great stuff.

Prompt: dark, cinematic, synth drums, slow build, ethereal pad, epic chord progression

[charred tendon]

Have you tried cat meowing, with an O?

[acoustic ridge]

that sounds really bad, lol

has anyone tried if it can do black metal yet? 😄

[charred tendon]

I did heavy metal....and yeah, no. lol LOTS of screeching and what sounded like distortion. Im sure I could have fine tuned the prompt, but just heavy metal alone was rough to hear. lol

[neat python]

yeah its horrid

[north dagger]

Hello

When can we download the open source version????

[unborn acorn]

that one was a little too ambitious in one go maybe 😅 prompt : "futuristic-jazz-with-808-kick-and-cow-bells-funky-guitar-riffs-and-energetic-vibraphone-solo"

[queen gust]

I can't access the pro tier features. The site knows that I had paid (in the billing section) but allow me to generate 90 sec tracks. If this was a zombie horde like in Days Gone, the Stable Audio team would be devoured to the point of no revival or restart. GAME OVER!

[novel frigate]

Are there any plans to make the creation of sound effects and short stuff <10 secs cheaper than maybe the full 90 seconds?

[tired halo]

any recommendation for TTS text to speech webui ?

I want to try out some of the facebook tts releases.

[harsh hatch]

Something happened. Anyone who wants can rate it

[acoustic ridge]

it can't do metal

[rough crest]

so, when does stableaudio come with their own infinite radio channel?

[queen gust]

I don't know about you but I'm getting sick of renaming the default downloaded filename. Your prompt = your downloaded file's name. My " idea ": Your prompt = downloaded filename* . * represents numbers 0 to 499. The site must never give you duplicate files.

[stoic jolt]

How do you run Stable Audio locally or in the cloud? Is there a guide?

[shut saddle]

No one is more disappointed in this than dadabots. Believe me. 🙂

[untold oasis]

output is clipping hard for some reason in tracks it generates for me

[civic rampart]

generation is fast today

[charred tendon]

Sounds like a cat with asthma. lol

[river latch]

I loved your gen so much, I decided to spend a few hours making something of it. Great stuff.

[tight anchor]

Love it! Very nice

[charred tendon]

I wish I can generate stuff. lol My account is still not connecting to the Stripe site where it shows I paid. Hopefully it gets resolved. But yeah, the Generate section of the site def loads quicker today than yesterday, so thats good at least

[true magnet]

Hey, is there anyone from the Stability team available in here? I have some technical questions about the new Audio generating tool.

@tight anchor you're on the team?

[copper glen]

Contact support, they manually adjusted limits on my account. By any chance, used a google linked account?

[tight anchor]

yeah I trained the model, wassup?

[charred tendon]

Contacted them already, and told them the issue. Still no resolution. I originally went through Google login, but then I manually created an account, and still there are issues. Both accounts, once I try to "Upgrade to Pro", once it gets to the Stripe site, it says I already have an account So yeah, the account part, is not "seeing" the Stripe payment section for me. And I provided proof via screenshots to the person who did reply

[true magnet]

Oh wow, really nice to meet you. Would you mind if I DM'ed you?

[copper glen]

Yes, same case for me.

[charred tendon]

Im sure it will get resolved soon; Just sucks, thats all. lol

[copper glen]

It's still showing upgrade, they just manually adjusted the limit

Yeah

[tight anchor]

Sure, I can give some technical info depending on the questions

[charred tendon]

Maybe they will do that for me at some point till it gets resolved. Can you do full 90 second generations and download them as WAV files?

[copper glen]

90 sec but mp3 only lol

[charred tendon]

Ah ok

[copper glen]

and gen is stuck again 🥲

[charred tendon]

Just tried it, and for me it worked. So it will probably happen in waves, depending on how many people are gen'ing it

[copper glen]

yeah, seems to happen randomly and then it's just sitting there lost for half an hour

generating on gpu /dev/null

[shut saddle]

He's head of Harmonai.

[copper glen]

bit janky at the second half but great start

prompted gpt to give me an example of an average track description on audiosparx

surprised it actually made something clean out of that gigantic copy paste

[lunar drum]

The theramin just isn't. Prompt: theramin solo, slow, mournful

Not sure what that is.

[river latch]

That's a *very * interesting sounding piano.

[next gyro]

🐉 🧙‍♂️ 🏰

[lunar drum]

Prompt: washboard solo, 240 bpm

It sounds more like a strange fiddle. The instruments are sure getting mixed up.

[shut saddle]

No 🐱 though...

[frosty wharf]

"They'll never know" -

[copper glen]

dropping an audio prompt into img gen gets you instant cd cover

[queen dagger]

This is

[frosty wharf]

[wind hedge]

My generation quality has noticeably gotten worse since day one

As in audio quality

[copper glen]

Do shorter tracks sound better?

[shut saddle]

I'd chalk that up to still learning how to prompt audio well.

[wind hedge]

Used same prompt and was different from yesterday

[copper glen]

pray for blessed rng

[wind hedge]

Yep

[copper glen]

a serene melody... played after three bottles of wine 🤣

[shut saddle]

Believe me, this is better text conditioning than it used to be.

It's still finicky but it's nowhere how it used to be.

[wind hedge]

Instead of "masterpiece" and "high quality" we are now gonna have "mastered" and "EQed"

[copper glen]

and microphone brands instead of camera models

hey that sounds like me playing

recorded on fisher price casette deck

[frosty wharf]

definitely struggling between two different keys at once, lol

[copper glen]

this is kinda nice

[frosty wharf]

F key, noice

[copper glen]

whimsical in any prompt really adds a couple bottles of wine

art

[frosty wharf]

burb

https://tenor.com/view/reeces-puff-dance-rasputin-just-dance-meme-rasputin-just-dance-tiktok-reeces-puffs-dance-reeces-dance-gif-20993325

[next gyro]

First variant I got to try was pretty funny to prompt heh. This is def a big jump for audio text conditioning with diffusion

[copper glen]

bit noisy

[tight anchor]

Theremin solo, slow, mournful|Instruments: Theremin

not the most melodic, but definitely more theremin

[tight anchor]

ahaha this rocks

[copper glen]

🧘‍♂️

[copper glen]

play at same time (start rhythmic when atmospheric hits 7sec)

[radiant stump]

Is there a way to makes sounds are sfx specific, keep getting "music" , typewriter was like weird guitar

[fallow sage]

that slaps

[small harbor]

If this would be normal quality audio, it would actually be good. Stable Audios audioquality is quite low for paid service. Maybe at some point there is Pro-PRO version of it.

Whatever drugs they are taking, can i have some?

It's probably ketamine. Stability Audios tape heads should be cleaned and alligned.

[timber garden]

Can’t download the file though I subscribe

[iron yew]

/fantastic ohm chanting

fantastic ohm

https://giphy.com/gifs/seth-jinzhan-gKr21u11F3hyRyJHqa

ohm chant

https://tenor.com/bEWLJ.gif

[charred tendon]

Just adding this here for anyone looking to have a base to build from when using Stable Audio. I asked this of Bing Chat and plugged some of them in myself, and they sound great:

1. E Major chord with a Standard 8th note groove layered with a Bass Guitar: This combination is common in rock and pop music.
2. A Major chord with a Four to the floor beat layered with a Synthesizer: This combination is often used in dance and electronic music.
3. C Major chord with a Shuffle groove layered with a Harmonica: This combination is great for blues and jazz.
4. G Major chord with a 16th note groove layered with an Electric Guitar: This combination is often used in funk and R&B.
5. D Major chord with a 12/8 groove layered with an Organ: This combination is common in blues and soul music.
6. A minor chord with a Motown groove layered with a Horn Section: This combination is characteristic of many classic Motown songs.
7. E minor chord with a Reggae groove layered with a Melodica: This combination is perfect for reggae music.
8. D minor chord with a Disco groove layered with a String Section: This combination is great for disco and dance music.

And like I said, just add things to these well known beats if you will, for those respective genres

For anyone looking for Metal suggestions, here are some starters:

1. E minor chord with a Double Bass Drumming layered with a Distorted Electric Guitar: This combination is common in traditional heavy metal.
2. D Major chord with a Blast Beat layered with a Growling Vocals: This combination is often used in death metal.
3. A minor chord with a D-beat layered with a Bass Guitar: This combination is great for punk-influenced metal genres like thrash metal.
4. C Major chord with a Breakdown layered with a Lead Guitar Solo: This combination is often used in metalcore.
5. G Major chord with a Gallop Rhythm layered with a Harmonized Guitar Riff: This combination is common in genres like power metal and NWOBHM (New Wave of British Heavy Metal).

[tight anchor]

Try adding in “TrackType: SFX” or the words “sound effect” or something

[small harbor]

Stable Audio vs MusicGen. MusicGen wins.

[fluid flume]

yes, when well the image to audio models be released?

[shut saddle]

Soon tm

[copper glen]

they're sniffing damaged casette tape powder

[fluid flume]

this is insanely impressive. I've been into stable diffusion image rendering for a while now, but I did not check out the audio side of things until today. another insanely addictive time sink for me, lol

so unsettling

[real skiff]

https://giphy.com/gifs/nbcgoodgirls-nbc-210-good-girls-dWNiglgPz5aKdVRG3B

How do I use the music generator?

[fluid flume]

link at the top of the screen

[shut saddle]

https://tenor.com/view/check-the-pins-kowalski-penguin-gif-20219688

[copper glen]

type out your prompt on your typewriter and fax it to stable audio

then your fax machine will play the song

[fallow solar]

strange - I cant hear music via webbrowser (safari, ipad ) but if I download the file - its fine.

[nocturne matrix]

Is the model out or nah?

[fluid flume]

when will stable audio controlnet be available?

[small harbor]

Same here. It's a bug as it used to work.

[storm thicket]

Any news on when an official API will be released?

[tropic obsidian]

If it 's a diffusion model we should be able to denoise an audio file and generate variations, right? Like img2img?
Does anyone know when / if it's being released?

[shut saddle]

That stable audio model is created using the audio library provided by AudioSparx. There is a revenue split that goes towards supporting AudioSparx and the artists that contributed to the training library. So that model isn't going to be released to the public.

The team is training an open source model that is not trained on that data, but on copyright free music. That model is still being developed and will be released later with the Harmonai tool suite that Fauno is developing. He doesn't have a set date when it will be released, just that it will be released when it's ready.

[cyan quest]

guys how do you do that

what is the command

[west veldt]

Quite interested in the denoising part and potentially frequency reconstruction.. I have a lot of tango music I use when DJing that would benefit from that 🙂

[hard steeple]

Trip Hop, Cinematic, soundtrack, Massive Attack, Bristol, Drum Kit, Shaker, Double Bass, Dramatic Reverbed Piano, Powerful Choir, Electric Guitars, Cool, Moody, Melancholic, Atmospheric, Dreamy, groovy, introspective, thoughtful, beautiful, Spacious, 90 BPM

I was QA and Audio Engineer during the stable audio project. If anyone needs help with prompting, the guide was based on my notes, and quite a few of the examples and most the default prompts are my creations. Would be more than happy to help anyone wanting to learn best prompt practices with the model!

[west veldt]

😆

[north plaza]

How to prompt for audio/music?

[autumn elm]

Is spectrogram used to represent the output of the model?

[small harbor]

How to create 90's rave stabs?

[copper glen]

lmao razor-sharp synth leads

[small harbor]

Since it's so easy to get bad results with SA, let's also share some ok generations/prompts. Prompt: Neo-classical, film score, Ludovico Einaudi, spacious, cello, violin, subtle percussions, grand piano, melancholic, breathtaking, elegant, crescendos, emotional peaks and valleys, 80BPM

Prompt: Vaporwave, Future Funk, mallsoft, '80s Japan, cassette, nostalgic, synths, drum machine, saxophone solo, city nights, neon, dreamy, aesthetic, relaxing, groovy, soft, 95BPM

[small harbor]

Stable Audio Feature request list:
SEED - Very important for tweaking prompts.
Multigenerartion 1-4 from one prompt.
Prompt sequence: 0-30s Promp1, 30-60s Prompt2, 60-90s prompt3.
audio2audio: Similar to img2img in SD.Drop in audio and get variations from it.
Remove the fade from the end that happens way too often.

[copper glen]

Tacking EDM Remix at the end of a tribal drumming prompt consistently gets these tighter modern rhythms

[wide blaze]

Hi all, can I use my generated AI audios with a Pro account to my Web Radio ? Is there some claim problems ?

[steady remnant]

prompts: hawaiian-melodic-psytrance-ethnic-with-mongolian-tribe-singing-mixed-with-berlin-dark-techno-with-strong-bass-turned-into-best-festival-set-on-tomorrowland_091823.mp3

Can I find a file or a list with used prompts for training the stable audio model?

[plush glen]

hey guys

[digital sparrow]

I think so.

[quasi drum]

This is like having session musicians do what you tell them without any complaints

[wide blaze]

So it seems to be good for webradios. Thanks a lot 🙂

[hard steeple]

going to have a look now for you! i'll see if I can make some rave stabs, it's probably down to the way your prompting by the looks of it! let me see!

[hard steeple]

90s Rave, Rave Stab, Chord, Sample, Synthersizer, Major Key, Short Sustain, Sharp Attack, Short Release, Nostalgic, Retro There we go if you just replace the synthersizer with whatever instrument you want or adjust this prompt to what you want this should work. Sounds pretty rave stab to me! The key to prompting this model is to supply it with as much context for it to find what your trying to do. It often helps if you already know the right musical terminology to decribe what you are looking for!

[hard steeple]

There is a definite for lack of a better term a prompt engineering skill gap for this model. In my personal opinion it might be a little to expectant of people knowing and visualising things in the right cohesive musical language. I'm sure this will get better and easier to prompt as we continue improving it.

[compact sand]

hi all. anyone has a promo code to Pro Subscription? thanks

[fair delta]

bro i literally can not wait for the open source model

[sly jetty]

good to know you can use keys/scales? do you know if specific keys works? like D minor, or Gsus4

[hard steeple]

If I can remember correctly maybe @tight anchor can confirm this. It was conditioned on key metadata, or at the very least it would of picked up some sort of semantic understanding from descriptions of songs. In my personal experience it's not exact, it works some of the time. hopefully this improves =]

[tight anchor]

There wasn't explicit key metadata, but it might know some stuff from descriptions

[sly jetty]

minor key works, arpeggios not so much

[fair delta]

freesound.org ?

whats the difference between stable audio and dance diffusion. is it just different models

[sly jetty]

love it so far, good work guys! this one has kind of an early trip-hop feeling

[tight anchor]

Yeah that's most of it

[tight anchor]

Dance diffusion is unconditional, no prompts

[shut saddle]

Because dance diffusion was unconditional, most of its use cases was in audio2audio "style transfer".

There were plenty of people who used it for sample generating though.

[hard steeple]

Nu-Disco, Detroit, Detroit House, Club Friendly, 90s, Vintage Drum Machines, Motown String Ensemble, Electric Guitars, Supporting Vocal Bed, Celebratory, Beautiful, Dancy, Climactic, Groovy, Funky, Dynamic, 115 BPM

https://tenor.com/view/pepe-headphones-music-gif-14789865

[shut saddle]

Daft Punk B sides over here.

[hard steeple]

Yeah mayne this sounds sick!

[frosty wharf]

"They will never know.......again, hehe" -

[jagged arch]

hi，my friends

Do you all use stable diffusion?

[rugged lagoon]

that's fabulous and right up my alley. I like the British band Broadcast is you know them. They are into sort of retro-futuristic, musique concrete, bbc telephonic workshop, sonic collages

Of course there's Trish Keenan's vocals but it can't emulate them

[rugged lagoon]

this is such a jam

this is so cool. I wanna make something sparse. Can I reference specific songs? I really like music with a fat beat, textured sounds, a little... funky

Can you reference specific songs?

[next gyro]

🌃

[rugged lagoon]

How do you prompt for dynamics, like what if I want the instruments to kinda drop out and there just to be a little beat but then they fade back in

I'm thinking like Radiohead's "You and Whose Army", or like a build-up

[tight anchor]

maybe try something like "Breakdown" in your prompt?

[rugged lagoon]

ok. can I reference songs explicitly? apparently so. Can I be like, "Unchained Melody" and "Mr. Postman" had a stillborn baby in C minor?

[tight anchor]

You can try, but it's not trained on popular music like that so it might not know the songs.

[rugged lagoon]

ah. are you sure? the best result I got I explicitly referenced a song. But you know how it is, noise is usually good within a limit

not two songs though

[tight anchor]

Some people in the dataset may say their song sound like another song, so it could get picked up from there, or be a coincidence

I tried to make bluegrass funk and I think I reinvented Blues Traveler

[rugged lagoon]

this stuff is too cool

[tight anchor]

Glad you like it 😄

[rugged lagoon]

haha I'm the king of finding patterns where there are none, it's my crummy superpower

I also recommend doing things like "revolver-era beatles"

like you know, bands change, etc.

[tight anchor]

Subgenre: Irish Drinking Polka|Polka drinking song with acoustic Irish instrumentation, groovy, Dancy, Full Mix, Grade: Featured, Grade: Featured, oompah, steady beat, 120 BPM, clapping

[rugged lagoon]

@tight anchor are you the architect?

[tight anchor]

Yeah, I trained the model

Subgenre: Irish Drinking Ballad|Irish drinking song with morose vocals, singing of a lost love, group vocals, sad melodies of a tearful loss, beautiful sadness, acoustic Irish instrumentation|Grade: Featured

[rugged lagoon]

Would it understand something like waltz-tempo? I like my stuff pretty sparse and scaled down, repetition is fine, like minimalist electronic music

[tight anchor]

sometimes, if you prompt it in the right way (still not sure what the right way is for most things)

I've seen it do 3/4, but it certainly has a 4/4 bias

[rugged lagoon]

I see. Yeah like you know how sometimes you'll prompt an AI art generator and it will generate the subject you're prompting for but there'll be like an echo subject because of the way you phrased it?

[tight anchor]

Genre: Classical|Subgenre: Waltz|3/4 time waltz music

[rugged lagoon]

also if I want something that has like an... untuned piano or something, discordant, do you know if that works? I think I'm running out of credits. Or do I get to do unlimited so long as I don't make them longer than a 1:30

I love that. That stilted playing, very nice

[tight anchor]

the credit count in the top-right corner should be accurate, you can subscribe to Pro to have up to 500 per month

which takes a long time to get through

[rugged lagoon]

I see. ok very cool

[tight anchor]

Genre: Piano|Subgenre: Ragtime Waltz|Classic ragtime waltz music, energetic piano chaotic haha

[rugged lagoon]

is there a way to tell if a token isn't really powerful? like if I use kalimba, would it know it?

[tight anchor]

best way to find out is to try it

it tends to know instruments pretty well

Genre: Piano|Subgenre: Pop|Modern pop piano music, catchy, simple melodies, pop drums, instrumental cover of that one popular song, emotional, dynamic

Genre: Piano|Subgenre: Emotional Meditation Ambient Piano

Genre: Piano|Subgenre: Emotional Ambient Piano Liquid DnB, soft breakbeats, 174 bpm, soothing, peaceful

[fierce night]

Will the open source model be able to create more than 90 seconds for one song?

[snow pier]

This is the best AI Audio Generation I have heard to date... Who ever trained this model should be congratulated! Amazing!

[snow pier]

Come On! How in God's Name is that possible. It sounds way to good to be true. I'm absolutely floored by this. I don't know about the rest of you however, I've been keeping a close eye and an close ear on the other attempts of Audio/Music Generation. Nothing even comes close.

[shut saddle]

Length isn't as limiting of a factor as image models tend to be. You do tend to get some weird quirks if you try to prompt a longer section of audio than what the model was trained on, but it still tends to be coherent.

[manic patrol]

i get on the website this "Oops... Expiration Time (exp) claim error in the ID token; ..." can some one help me out what s wrong?

[shut saddle]

That just sounds like it timed out.

[manic patrol]

but i want be get timed in

🙂

[shut saddle]

I'd just try refreshing your browser and trying again.

[manic patrol]

tried that since this morning

even with inkongito and google login it doesnt work

[hard steeple]

round of applause for @tight anchor @errant cosmos =d their the architects. =] I just QA'd it

[warped mist]

genre

[hard steeple]

it's not a bot haha https://www.stableaudio.com/ go here!

[wide blaze]

Hi, is there a way to ask for features, like Folders for generated music ?

[manic patrol]

where to post bugreports?

contact link is dead

[shut saddle]

Try posting this in the troubleshooting section of Harmonai.

Disregard this.

[manic patrol]

ty

[tight anchor]

For bugs with the service, please contact [e-mail]

[shut saddle]

Maybe that needs a pin?

[tiny frigate]

Weeeeeeeee

[slender matrix]

is it open source yet or still paywalled royalty free generative stuff

[quiet delta]

hmm, my first result was very discordant :/

[fossil adder]

im getting very discordant stuff too

its also so LOUD

[slender matrix]

paying doesnt make it better either btw but you still dont have unlimited access

it feels

cashgrabby almost

[fossil adder]

I like this vibe but I wish it had more diversity it is just repeating

[slender matrix]

like pushing experimental tech to market before people even made it worth using

i believe in local AI more than saas

[tight anchor]

Working on the open-source versions 🙂

[silent lintel]

do yall think that this model will run in any low end pcs? seems a lot cheaper than image generation

[slender matrix]

maybe im too critical

[fossil adder]

need variation and 'inpainting/inmusicing?' features 😛

[tight anchor]

The model can be hit and miss, learning how to prompt the model properly can certainly improve outputs

[slender matrix]

but you just see it everywhere in the ai space

[tight anchor]

also make sure to hit the thumbs down on ones you don't like, and thumbs up the ones you do

[bleak wigeon]

it's ok

[hushed horizon]

It was 1B parameters roughly iirc?

[fossil adder]

this is very cool though ive been looking for something like this

[hushed horizon]

so that should run easily even on low-end cards

[silent lintel]

nice

[slender matrix]

audio gen is very difficult to pull off because of phasing and the amount of information you have to condense

[fossil adder]

i feel like most people that use ai don't know how to use docker or set up stuff locally haha

[hushed horizon]

Yeah for sure I'm waiting for the local version

[silent lintel]

we'll see

[slender matrix]

And thats why we have Gatekeeping

[slender matrix]

Yea I hope interesting ideas can be pulled off

[fossil adder]

hey im all for releasing it, but like, ive seen the amount of people that fail to just set up sd locally <_>

[abstract lantern]

beautiful, beautiful, this is more better than the mid musicLM

[abstract lantern]

no way guy of rvc is here

[slender matrix]

[shy mango]

any good prompts for lofi hip hop? so far most of the outputs i get are off pitch

[abstract lantern]

hows it going guy rvc

[hushed horizon]

No offense but Beethoven would never write that cacophony :P

[silent lintel]

i bet these audio models are gonna get insane in 1-2 years

[abstract lantern]

sounds like one of beethovens scraps

[hushed horizon]

It really falls apart towards the end

like 0:30 and onwards

Just slamming the keyboard randomly :P

[silent lintel]

in my case i think it got better but bad at the start

[hushed horizon]

I mean the classical piece to be clear

well, "classical" :P

[barren parrot]

this is nuts, but i need a negative prompt 😭

[hushed horizon]

One of the things I'm very curious about is if it can do classical-like compositions as opposed to just nice beats and repetitive/textural stuff

[barren parrot]

is this gonna be released open source?

[abstract lantern]

negative prompting and stable audio v2

thats what i want

[edgy hound]

are they offering any trial other than the 20 tracks?

[shut saddle]

There will be open source tools released but they aren't ready yet.

[abstract lantern]

hol on

trials?

srs?

[shut saddle]

You get 20 free gens per month.

[abstract lantern]

so its basically like novelai

can gens overfill?

[edgy hound]

its only 20 test which you will consume it with less than 20 min of playing around.

[shut saddle]

This^

[abstract lantern]

mm, sounds like keys

[tardy patrol]

if you think about it coin is the key to the doors of this world

[hushed horizon]

How much of the classical corpus is public domain

Does it all like belong to Disney or something

Although I guess for training this takes recordings of performances, right? And not just notes.

[bleak wigeon]

i've only generated one thing that is actually listenable

[hushed horizon]

There might not be that much in the public domain for classical pieces

[silent lintel]

funny how there are random explosion sounds lol

[barren parrot]

I need to add something like (disonance:1.7) to my negative prompt 😂

[abstract lantern]

also is deepfloyd emerging as a lora soon?

[fiery robin]

i just put the prompt "screaming" in stable audio and its nightmare fuel

[tardy patrol]

yo the first 3 seconds goes hard tho

that last bit at the 3rd second is actually a very nice crunchy snare

[barren parrot]

we're gonna bring back michael jackson without his love for boys this time

[abstract lantern]

this audio gives me hooded kars vibes

[silent lintel]

its not supposed to be a song tho

[tardy patrol]

i know im just point out the crunchiness

[shut saddle]

Audio prompting currently is vibes based instead of precise control. Getting your audio in the right direction is tricky enough.

[tardy patrol]

very sample-able

[abstract lantern]

quite a di molto moment

[shut saddle]

Just from my personal experience prompting for audio.

[abstract lantern]

i love this

[hushed horizon]

that one kind of works

[tardy patrol]

thats nice

its pretty good at texture and vibe it seems, like Geck said

[shut saddle]

I just think that we don't often have the right vocabulary to describe audio.

[rare oxide]

wacky

[tardy patrol]

lmfao the bass got sent back to the 1880s

[silent lintel]

damn this turned out really well

[rare oxide]

tbh sounds like a fairly nice song except i am locked in a rubber room

[abstract lantern]

crazy

with rats, right?

hopefully you recover

this takes the cake

[rare oxide]

payday 2 soundtrack

[tardy patrol]

yeah i agree. its like, there's always a lotta buzzwords in audio engineering and music circles, like 'crisp', 'punchy', 'bright', 'flat' etc, which dont always make the most immediate, intuitive sense

obviously if you do that kind of stuff, you share the vocab, but for images and art the vocab is more concrete i think

[rare oxide]

holy heavens

[abstract lantern]

more breakcore than synthwave tbh

[rare oxide]

0:24 slaps

[bleak wigeon]

yeah i like that part

[silent lintel]

just needs a little bit less weirdness

[random galleon]

is there going to be a local version of this too?

[abstract lantern]

when your child plays with the fax machine:

[rare oxide]

truly the 8 bit chiptune of all time

[bleak wigeon]

splatoon character goes through the printer

[abstract lantern]

https://tenor.com/view/splatoon-marina-off-the-hook-picmix-gif-4924056445864584123

[abstract lantern]

breakcore is quite well

[tardy patrol]

lmfao this reminds me of death grips - culture shock

the printer part at least

[silent lintel]

did it try making breakcore with train sounds lmao

[abstract lantern]

ye

the beginning is gud

[tardy patrol]

goddammit all i got was duds

not a single prompt worked for me

see you next month

[silent lintel]

the problem with using authors as reference, is that the model probably doesnt know any

[tardy patrol]

yeah i tried with and without

no lucky

🥶

[abstract lantern]

audio to audio when

[barren parrot]

this shit has so much potential

[tardy patrol]

i havent heard a single vocal in any of the things i generated or the ones i listened to here

is that dataset all instrumental?

[tardy patrol]

couple of months i say

mayhaps

[silent lintel]

does this count as one?

it turned out like catacombs kinda thing

[inland bear]

I showed this off to someone, they put my discord username and came out this 💀

[tardy patrol]

yeah, thats pretty nice, sorta creepy

[silent lintel]

the ai is tryna communicate 💀

[tardy patrol]

https://tenor.com/view/better-call-saul-chuck-mcgill-chicanery-breakdown-gif-25422432

[inland bear]

tried to make some ambience for some game to test it out and it works well

[silent lintel]

thats pretty good

[tardy patrol]

Ambient Diffusion

[silent lintel]

oh i havent heard jazz yet

[shut saddle]

https://tenor.com/view/better-call-saul-saul-goodman-saul-ok-gif-22735393

[tidal lava]

are there stable audio models I can download?

[shut saddle]

They haven't released the stable audio model and probably won't since there's a revenue share split with AudioSparx and the artists that contributed to that model.

[silent lintel]

the open source one isn't available for testing yet?

[shut saddle]

The open source model isn't ready.

Fauno also wants to release the tool suite at the same time, so we likely won't have anything until both are ready.

I look at stable audio as a proof of concept.

[icy phoenix]

this had potential til the 12 second mark

[lyric finch]

I tried it out until I ran out of free tracks. I found a lot of distortion in my early efforts, and the tracks I tried to be thematic didn't turn out to be very melodic.

[tight anchor]

It's mostly instrumental, we currently add to the prompts in Stable Audio to bias it towards instrumentals because cursed vocals kept popping up in outputs unprompted

[lyric finch]

On interesting track I made was simple a tuba riff. Interstingly, there was a point I heard the tubist inhale, which I quite liked.

[tardy patrol]

yeah that makes sense, so legible human voice will be the "generate a human face/hands" barrier of the audio side of generation it seems like ?

[shut saddle]

There's several non stable audio programs that have made realistic audio. This model didn't specialize in it.

[rare wind]

YES! FINALLY! I have been waiting for something like this!

Funny... just a few hours ago (after months of not checking) I was looking around and found something similar... guess I'll have to test both.

A music is finally here... let's see if it becomes a trend.

[nimble torrent]

first shot out the gate too?

it's either pretty darn good or has a very consant level of decency regardless of your input

here's to hopin it's the former

[icy phoenix]

I used up literally all my free generations to make what I wanted to make sound good

[coarse lantern]

Now i really can't wait until stableaudio gets a git and custom models as well! :D

[hard steeple]

A personal fav output when I was working on the examples on the website that never got used. 😦

[nimble torrent]

trip hop, i like that sound of that

[hard steeple]

🙌

[nimble torrent]

for my second gen i tried a trap beat

lets just say i had a bad time

[hard steeple]

lmao there is loads of trap in the dataset and I have had mixed results it's really wierd. I swear it's a prompt away though!

[nimble torrent]

it's like the model kept shtting itself or something

[hard steeple]

it sounds like taking a hammer to the piano and tearing it apart hehe

what was your prompt

[nimble torrent]

Not really pleasing to the ear

Trap beat, hip hop, 808 drum machine, 808 kick, claps, shaker, piano melody, phonk samples, aggressive high hat patterns, underground rap

[hard steeple]

maybe try adding moods to your prompt, like how you want it to sound emotionally

that might add the context for the instrumentation to work off

[nimble torrent]

...... urban

[hard steeple]

oh dear

nooooooooooooooooooooo

[nimble torrent]

[hard steeple]

stuff like, gritty, raw, moody etc

[nimble torrent]

ok ty ill give her a go

[hard steeple]

like I said i've had mixed results with trap but there must be a way i refuse to believe there isn't

[shut saddle]

Any chance we can get a token list of frequently used words that were used in the training dataset? 🙂

[nimble torrent]

those are really my only 2 lanes unfortuantely

disgusting heavy death metal, hip hop trap

[shut saddle]

If not, that's ok. The discovery process is fun too.

[hard steeple]

you can literally pull up the audiosparx website and take a look yourself

[shut saddle]

Good to know! Thanks!

[hard steeple]

I did that to figure it out

[mellow cosmos]

where is the model? is it not open source?

[nimble torrent]

https://stability.ai/research/stable-audio-efficient-timing-latent-diffusion

[hard steeple]

this model is trained on licenced data! we are currently working on open source models utilizing what's available for us to use

[nimble torrent]

"Our first audio AI model is exclusively trained on music provided by AudioSparx, a leading music library and stock audio website."

[mellow cosmos]

no DL there. am I not seeing something?

or is this not meant to be run locally like SDXL

[nimble torrent]

that's all that's public rn

[hard steeple]

it's a web ui https://www.stableaudio.com/

[mellow cosmos]

you knew what I meant, I thought it was ready for local use

[nimble torrent]

can only use free stuff in your own production, paid accounts can use commercially, nobody can use output to train a model

[mellow cosmos]

I was talking about the weight themselves, I just thought this is going to be open sourced like the rest of SAI stuff

[hard steeple]

you wouldn't be able to run this locally unless you have the compute. it's to big to run locally, currently, we are making an open source model that you will be able to

[nimble torrent]

Right, but I gave you all that's available on it rn

[mellow cosmos]

if requirements are like SDXL it shouldn't be an issue

[nimble torrent]

yep haha no issue at all right guys

[mellow cosmos]

? with stuff like AIT, SDXL can run at 12it/s on normal GPUs

[nimble torrent]

define normal

[mellow cosmos]

4000 series, 3000 series

[nimble torrent]

I render animations at 1it/s on my 2060 6GB

[tight anchor]

I mean it would run locally, but we're working on open source versions yeah

should be fine-tunable on your own computer

particularly as people find out things like LoRAs

[shut saddle]

This is a completely unrelated model. It's not SDXL.

[nimble torrent]

they were comparing the requirements not the models

[spice crow]

i love it

[mellow cosmos]

though when blending images with SDXL the speeds can shift sometimes

[nimble torrent]

oh it does vocals?

I didn even realize

well

"vocals"

[spice crow]

hahahah

[shut saddle]

It can, it's just not great at it.

[nimble torrent]

GtawaAAAAAAAAAAAEYYYEEEEaataweTWeaaayEEEEEEEYEAW

[spice crow]

are u related to 100 geccs

[nimble torrent]

when she said that, I felt it

[mellow cosmos]

will it be able to blend audios like SDXL can blend images? or is that aiming too close to the sun

[tight anchor]

We'd probably need to add some audio conditioning to the model for it to do that properly

[mellow cosmos]

and stuff like AIT and OneFlow will be compatible, right?

[hushed horizon]

quality

it's surprisingly musical

[hard steeple]

hahahaha that's sick

Wild West, High Noon Shoot Out, Cinematic, Soundtrack, ennio morricone, Percussion, Whistles, Horses, Action Scene, SFX, Shaker, Guitar, Bass, Timpani, Strings, Tense, Climactic, Atmospheric, Moody, BPM 80

https://tenor.com/view/the-office-gif-19830161

[spice crow]

i'm finding it tough to make anything listenable 😅 I am rlly trying to make it do consistent vocals tho, vocals that dont suddenly cut in and out

is anyone else getting a lot of microphone-feedback like quality?

[hushed horizon]

Whelp, that's my 20 generations

I think it managed to spit out an reasonable piano piece at least?

Well, it cuts off early, but what's here makes some sense

want my money back for those last 5 second tho

[abstract lantern]

audio2audio is gonna be sick

perhaps like img2img

with a text prompt

[vague cliff]

Tried a few things. Overall the sound quality is much better than previous models I've played with, so it's a step in the right direction, but I'm not sure I could get much that's really usable out of it given the current state of it. I'd say the licensing terms are ambitious in relation to the current quality. Maybe you have different markets in mind than the typical home studio enthusiast, but anything that steps too far beyond the familiar terms for royalty-free samples is going to be a hard sell. That being said, the jump in quality is exciting, and I look forward to checking out how things develop (and reading the papers!).

[spice crow]

how do u get better sound quality

this is my favorite from using up my 20 generations haha

[vague cliff]

I'm not saying the sound quality is good, I'm just saying it's better than other generative models I've tried to date. The first generative model I tried was google's (years ago), and they did everything by squashing the audio with mu-law compression (IIRC), so everything it generated sounded like it had been smashed into oblivion. Most generative audio models seem to have followed that pattern. This is one of the first things I've heard that doesn't sound like everything is compressed to death.

[tight anchor]

Yeah we use a learned autoencoder to compress the audio to death, sometimes the decoder properly brings it back to life haha

[viral fractal]

It needs to be as easy as the rest of them (google and other opensource ones) if they want to charge us (I know you don't have to).

For right now it sounds like shit.

[tight anchor]

It's a tradeoff between high compression allowing longer sequence lengths, but also making the artifacts worse

I'm certainly looking for ways to make it sound better

[viral fractal]

I can wait though. It does feel unusable and hyped (by accident)

[tight anchor]

It sounds great for some outputs and styles, and abysmal for others

[sharp crow]

any easy to install local webui yet? for 6Gb vram?

[spice crow]

i love it, keep up the good work 🙏

[viral fractal]

is the database just poorly sorted or too large?

[spice crow]

soon my vocaloid garage punk band will be ready to world tour

[tight anchor]

The dataset is great, there's just a lot of technical stuff to get it to sound good, be quick, and work with long sequence lengths

And right now in particular the quality can vary greatly depending on how you prompt

a lot of the dataset is sound effects and field recordings, so there's lots of non-musical stuff that gets picked up as well, which can probably make the music sound worse if it's not fully aware it should be making music

[sharp crow]

@tight anchor You wouldn't know any? also heya long time no see

[tight anchor]

Hey! No local web UI yet, but we're working on that for the open-source models

[viral fractal]

why is it fucking hard to use? why do I want to support the companies that I don't like? Why can't I just write something simple for it to follow?

Are the other ones just scamming the user? Are they pretending to understand the full prompt? Is this trying to understand the full prompt? Does it actually truly understand a large amount of what I type?

[sharp crow]

nice I asume it will be announced here?

[vague cliff]

One of the things that I tried to do a long time ago (when I was trying to use RNNs to generate audio) was to use vector quantization instead (instead of what was essentially mu-law + 8-bit scalar compression). It still introduces artifacts, but it's got an interesting character to it that can sound good in a synthesis context (e.g. I suspect that some of Roland's samplers use VQ internally--that's just speculation on my part though...).

[tight anchor]

Yes, though the Harmonai community is the main focus for our open source stuff

[sharp crow]

well I am member there as well

[tight anchor]

Yeah, RVQ-AEs are quite popular in this space, they can work well to make the latents "snap to grid" after the diffusion process

I have yet to train a good one that has the properties I want for diffusion, but it's certainly on my list of things to keep trying

[spice crow]

may u pls share an invite for Harmonai :o

[tight anchor]

https://discord.gg/9gG5Ezbj

[vague cliff]

Nice. Yeah, I've been so deep in the text-to-image literature that I haven't even tried to keep with the audio stuff. I knew people were using VQ in some capacity for audio now, but only in a handwavy way.

[sharp crow]

btw can't share them it get blocked shouldn't it be in the whitelist?

[tight anchor]

not sure why you can't share it, weird

[spice crow]

thank u 🙏

[tight anchor]

Most modern audio generation systems use residual vector quantizers as audio tokenizers to get token indices to feed into a transformer, treating the whole thing as an autoregressive language modeling problem

That has its pros and cons. Because you're generating the sequence one token at a time, the full power of the model is going in to "what is the next token?" That lends itself much better to generating melodies compared to diffusion where the whole sequence is being denoised at once

The tradeoff is that the inference time scales linearly with output length since you're going one token at a time

But also you can go as long as you want

Of course transformers will only have so much of a context window, so the model "forgets" the past at a certain point as it goes along

With diffusion it's the opposite. You've got the whole window all at once, so everything can see everything else. But you're also stuck generating the full window size for every output

[sly jetty]

the autobot blocks all discord.gg short addresses from users, they are mostly spam/scammers

[sharp crow]

Stable Diffusion discord links are whitelisted afaik. which they normally are but this seems to be not the case with harmonai

[tight anchor]

yeah, I also generated a new link so I see how it could be tricky, not sure how the whitelist feature works

[sly jetty]

I'll ask but yes, should be on a whitelist

[quiet delta]

probably my best sounding generation so far

[tight anchor]

This is also why the ask for longer output lengths is a bit non-trivial with diffusion. With transformers, it's just a matter of running the model longer. With diffusion, you either need to fine-tune the model to work with a longer length, or use outpainting to do "continuations"

[vague cliff]

Here's a thought: The primary use case I have for audio generation is sample/sound effect generation, so a fixed-length generator that sounded better would be vastly preferable to an arbitrary-length one. So even it if had the same artifacts as convolutional systems like SD (repetition/lack of coherence when the image size exceeds the primary training size), it would still be very usable for post-production music.

[hushed horizon]

What length are you targeting for the open source one?

[tight anchor]

Yup, for the open-source models, I'm more focused on short lengths (12s right now) with more of a focus on quality (...as much as FreeSound will let me)

that's mostly because it's more practical to fine-tune on smaller GPUs, and also because we just have way more sound effects to train on than songs for an open model

[vague cliff]

You're using the "free-to-use sounds" dataset for that as well right? I know I looked at a list of open datasets in the ElutherAI discord at one point.

[tight anchor]

I haven't pulled that in yet, but I may at some point.

The vast majority of the data for our open models right now is CC0/CC-BY data from FreeSound

[quiet delta]

im imagining some kind of hybrid approach. Starting with a song format like Verse-Chorus Structure/Head, solos, bridge for jazz/verses hook, bridge, break for hip hop, and maybe generating a short melody, then make variation segments to slot into each of those structure parts.

I feel like trying to make a long coherent piece will end up rambling unless we prod it towards a little bit of structure

[tight anchor]

Yeah, that would be awesome. If I had a well-labeled and segmented dataset like that, I'd love to have structure conditioning

[vague cliff]

It's really fantastic. Meticulous metadata. But yeah, any CC0 sources are ideal.

[shut saddle]

I'd imagine that once an open source model is out, CLAP can be better trained to tag existing audio with the appropriate details which will help improve future open source models.

[untold saddle]

I'm super excited to see this evolve

[terse narwhal]

is there going to be a local install version of stableaudio or on the website only?

[pliant chasm]

Check out Harmonai's Discord!

https://discord.gg/NAkjPfYqbJ

[distant ember]

Best of my 20.. not too bad!

[civic heron]

anyone know if this works with phonk/drift phonk style beats? can it do lyrics (memphis tapes, phonk tapes etc.)

[unique mason]

Wen API for this

[distant ember]

Sorry I'm not really familiar with the genre to make a better prompt. I'm not sure this is what you were hoping for lol

In general it doesn't seem to do lyrics

[civic rampart]

banger

[distant ember]

Aww yis

[fluid tulip]

voyage to mars, classical, deep bass (moon knight scene?)

The Truth and Goodness win against evil, epical, classical, melodic harmonical angelical choir, deep bass (Avatar Vibes...?)

voices of nature, flora and fauna made a classical orchestra, epic, deep bass and sound effects. (my opinion: good prompt, bad result)

Those are my 3 best shots, i think that until version 4.0, i personally will not enjoy it so much...

Starts as Epic deep bass scene and ends in arabic ballade (i enjoyed this one)

Deep hypnotic trance with Mongolian throat singing (like the first version of a Sci Fi scene recorded in the worst TS format ever, and OMG where is the Mongolian Throat voice that i so hoped to ear?)

[next forge]

synthwave but everything is slightly off key

[tidal lava]

i thought this came out nice (i do not know anything about music production and prompt it like it's sd, no claps? clapless no snares? snareless, idk)

[untold meadow]

what's the expected VRAM requirement and is it possible to create music that loops seemlessly? Excited to see what comes out of this, I have been waiting to have an AI generator for music to use as background music in games for a while now.

[grand matrix]

Looping seemlessly: it loops, but you can't tell when

[distant ember]

The way this one starts is 👌👌👌

[brazen flicker]

Is there a support room or ticket system?

[pliant chasm]

Support is currently handled through email: [e-mail]

[brazen flicker]

Thx! Great stuff so far, ai generated horror audio drama made from my phone while I charged my car. The future is now

[tidal lava]

generiere eine silberfarbene Roboterhand mit dunkelblauen Hintergrund

[junior mauve]

any idea when the open source release is coming?

[shut saddle]

If you have that answer, let me know next week's winning lotto numbers.

[tidal lava]

When you enter "hardcore music" into the prompt, it gets censored for some reason

[hard steeple]

working on it!

1970's Vintage RnB, Chicago, Isaac Hayes, Curtis Mayfield, Drum Kit, Rim Shot, Tambourine, Bongos, Brass Section, String Section, Rhodes Piano, Bass, Flute, Chimes, Groovy, Retro, Moody, Peaceful, Introspective, Thoughtful, Dynamic, Slow-Tempo, Sensual, Soothing, Relaxing, Spacious, 100 BPM

[tidal lava]

Drum and Bass, 90s Style, Upbeat, Energetic, Rhythmic, Angelic Female Vocals, Mastered, Synthesized Bassline, Jungle Breakbeats, Atmospheric Pads, Melodic Hooks, Dancefloor Vibes, Nostalgic, 170 BPM

[hard steeple]

that fill where it just wierdly devolves into spectral garbling kind of hard ngl

[nimble torrent]

You ever figure it out?

[nocturne zinc]

Yea nice buildup there, overall track is great and flows just fine, you caught a nice one!

[nocturne zinc]

I like this one, esp that hot part near the end that sounds like tape saturation.

[rugged lagoon]

Hey y'all, I hope you don't mind but I wanted to give some feedback.

This platform is very cool. If I'm honest, I think it sounds better when it's shorter

my favorite outputs have been like a cool intro or buildup to a song, like the beginning of "pictures of you" by the cure. Or an outro

Is there a way I can prompt for that? I guess I can try. But I'd be curious to see what it's like training the data set on song intros and outros only

[rugged lagoon]

just checked tha tout, that style of music sounds great!

[rugged lagoon]

I think that actually sounds pretty good, I just wish the vocals were a little slower

[rugged lagoon]

Does anyone know if there are any platforms that offer sound genre prompts for short clips of sound, like if it would give you suggestions of what that style is

[crystal cloud]

With some trepidation 🙂 Here's a range of results from my 20's worth. Exploring prompt structure. IMHO, overall pretty good start from StableAudio

[pure aurora]

there is something from me!

[sly jetty]

Somehow this prompt produced very glitchy distorted hi-pitched things, I was trying to aim the opposite way: "The soothing sounds of gentle ocean waves and rustling palm leaves accompanied by calming instrumental melodies for a blissful, serene experience."

A cool experiment I saw earlier in a tweet and expanded upon is describing what you want to an LLM (chatGPT, Claude, Beluga) for them to craft the prompt for stable audio.

Example1: I am in a zen state, on a tropical paradise, completely relaxed and in tune with nature. I like music. Describe what I should listen to in a short sentence as a list of descriptive words separated by commas.

Example2 (from Nao Tokui): You are a musicology professor. Describe the most obscure music you know in a short sentence as a list of descriptive words separated by commas.

[fluid tulip]

I don't sure if the right word can be used without being banned, so, id like a "Deutsche" Macarena.

[fluid tulip]

it sounds like voices, OMG can it be cacophony? maybe somebody can test that?

[pure aurora]

These are sounds from the damned who were forced to do their job

[sly jetty]

Great insights on that stage @errant cosmos and @hard steeple ! 👏

[shut saddle]

Good office hours session.

Also some awesome news about the progress on the open source model, but you had to be there to hear it. 👀

[nocturne zinc]

Howdy, I created two songs using: downtempo, minimal, 90 bpm, epic intro, synth, chopped, melodic, then I arranged them in Ableton to be more structured, added a lil bit of reverb and delay to smooth some things, and a big of mastering to make them sound more similar to each other for this:

This is the album art in the MP3, obviously also AI generative:

[nocturne zinc]

Huh trippy, you ended up with a decent ASMR track tho!

[nocturne zinc]

For that last one you got a nice piano piece, I have tried several times and it's like music to guide you into insanity, such dissonance.

[sly jetty]

If you guys want to try more moods/cinematic stuff, since the training was done with Audiosparx data, there is some metadata that produces good results. You can say stuff like "horror macabre" or "scifi robots" or "psychological thriller". Audiosparx is used in movies/tv so it has a lot of that inside, those genres and moods

[nocturne zinc]

oh interesting I have several tracks up on Audiosparx, maybe I'll get my own vibe in there like a oroborous.

[sly jetty]

that'll be a good experiment if you remember your metadata info you could probably get some of that back

[distant ember]

I live for the dissonant breakouts into melodic harmony in a lot of these generations. Maybe I'm a sucker for that sound

[nocturne zinc]

I really like the way some phrases are very distorted like tape saturation, very cool vibe.

[sly jetty]

if I remember correctly they are pulling BPM, moods, instruments and topics

and style of course

[nocturne zinc]

BPM was spot on for me, it was super easy to bring into Ableton.

very nice, i've used Riffusion before and it's also fun, but I have to manually set the tempo and length.

This was a bit too dissonant for my head, @distant ember

[sly jetty]

keys don't work, i specifically asked them about it, it does not understand keys

[nocturne zinc]

aha good to know!

[sly jetty]

you can prompt "minor" or "major"

but not specific keys

[distant ember]

Oh yeah there's very little coherent meat to that one. Just a pinch of dissonance is when it is best

[hard steeple]

TY!

[nocturne zinc]

Nice idea, trying this now!

Pretty cool, I used all 'mood' tags for one track: anguished, cold, contemplative, dark, defeated, depressed, disillusioned, emotional, heartbroken, helpless, hopeless, hurt, lost, moody, oppressed, painful, reflective, slow, sparse, suffocated

[sly jetty]

nice

someone did "cursed" 10 times with commas, amazing result

[shut saddle]

We had a few fails too but that's how it goes sometimes. Turns out "Medical Drama" just produces dialogue lol.

[distant ember]

To confirm, no time signature specificity? Asked for a few in 7/8 that seem to be in 4/4. Not that I need it, just playing and testing.

[sly jetty]

time signatures don't work

broad tags: style, mood, instrument, topic, regions

topics are things like: birds, addiction, birthday...

regions: latin, asian, specific countries and specific regions

you can also do genres, subgenres and BPM is a must

all movie/cinematic genres: blaxploitation, action, horror, scifi, western (and subgenres)

[shut saddle]

You could coax time signatures indirectly through genres.

[sly jetty]

true, waltz will tend to be 3/4

jazz has odd timmings

but you cant specify, I burnt like 10 generations trying no avail. 😂

[nocturne zinc]

First half is great but the rest just doesn't work.

[tight anchor]

We didn't include regions in the training prompts

but it might still work because the info is often in the description as well

and lots of regions/countries are represented as their own genres

a hell-bound ship's crew singing a sea shanty on their way to the depths, stomping their feet to the beat|Instruments: Male Voice, Ship Horn, The Wails of the Damned, Stomping Feet|44kHz stereo, high quality recording, steady stomping beat

[shut saddle]

Was there any token alterations made to the initial dataset or was the model trained off of the AudioSparx set as is? If it's the latter, the website does sort music by country.

[crimson turtle]

As a Nashville resident I attest that SA nailed this one. If you were hanging out in the honky-tonks on lower Broadway, you'd hear something like this: @tight anchor

[tight anchor]

Just a few hours ago I was thinking about when we checked out the honky-tonks on Broadway together and saw the guy playing pedal steel guitar, that was such an awesome experience!

[crimson turtle]

Hey how do we get a "sharing url", to post to social media, etc?

[tight anchor]

also eating some Nashville hot chicken rn haha

[tight anchor]

We're working on an improved share experience

I've seen people posting videos, honestly I'm not sure how they're doing it, probably some screen capture thing