📚┃suno-school

[south kindle]

its a download link with extra steps.

[sonic rune]

I don't see the download-video option in Discord… maybe in mobile, but I haven't tried it

[south kindle]

copy the content link to the cdn

the platform is still young and is more focused on making competitive features to keep up with the competition, so there are these rough edges we need to deal with for the time being.

I mean, we can set images for our playlists, but in most cases it still displays whatever the art is for the first track in the list.

[foggy salmon]

Hello folks! I found an interesting problem in the latest version. Yes, the sound quality has improved, and the duration of generation is pleasing. However, it seems to me that the developers have lowered the temperature too much, the performance has become more bland and less inventive. In addition, this version is more likely to ignore the style instructions and now and then goes towards pop, hip hop or bardic. Has anyone else encountered this or is it just my personal impression?

[sonic rune]

At least they make the older versions still usable, unlike several AI tools I've used that retire their old engines over time.

they do specify that the newest model is "better song structure" which means that it will be more limited in certain capacities, while the previous model boasts being "broad" and "versatile" which I think gets at what you're saying. Essentially, you're pointing out that the model is doing exactly what they said it would do.

[narrow cypress]

y'all...

in 3.5, in the lyrics field, this really works. Enter this at the top of the field:

then for each item to weigh, example:
[Softly: 3]

[languid spear]

In 3.5 no matter what it keeps generating 2 minute chunks, I'm trying to extend a song by adding a chorus and just an outro, but it just keeps on going, what is the way to tell the generation to stop?

[foggy salmon]

I suppose so, however, it doesn't explain that the new model is so limited considering the genre and style

[bronze helm]

I will try this actually

[narrow cypress]

more experimentation in #💬┃general-chat , may have disproven it after all, but definitely test it/modify it/change it and let me know if you have any success

[bronze helm]

I have a very small sample of tests (and little time at the moment) but it seems using just [Softly] is more successful than using [Softly: 3] for example (one generation ignored the Softly: 3 as a whole, while all generations using just Softly didn't ignore it)

Still, I think this is worth mentioning in #💡┃suggestions-web as prompt weights could be a pretty powerful feature

I'm waiting for the year where I can do mind2audio with perfect reconstruction /j

[cerulean flower]

Hi everyone !
Is it possible to make a song with 2 singers (1 male 1 female) ? My last attempts were complete disasters ahah

Wondering if someone succeeded 😁

Thanks !

[narrow cypress]

Your best bet would likely be including a style in the style field that typically is performed by a male and female, like "duet"

I haven't tried in 3.5 but in 3, I struggled with this a lot too

[cerulean flower]

yeah i was trying this too, I'll try again in 3.5 thanks 🙂

[quasi flower]

I often find a voice that I really like and then in another variation, I like the instrumentation. In the future, it would be great if it was possible to have the "singer" from one song perform another song,.

[steady rover]

Does the AI currently have the capability to make songs similar to Type O Negative?

[spare stump]

Hi.. can anyone help me use up remaining 5 credits. Can't get rid of them and I am afraid the 50 won't refresh for tomorrow

[prisma wraith]

Don't worry. you'll always be set back to 50 every day, unless you buy a subscription, which gives the months worth of daily credits up front

[grim elk]

Important question! .... Reeeeeally important...! Can Suno Clear it's throat?

[topaz pine]

hello

what is the problem? how can i fix it?

[prisma wraith]

Chirp's gone, please make use of the website instead https://suno.com/

[spare stump]

right.. but u don't get the refill of 50 unless u use up all your credits

[prisma wraith]

You do. but if you really want to use up the last 5 credits, if you have a multiple part song, you can get the whole song, it should cost you 5 credits

[spare stump]

good idea..i will try that

[topaz pine]

thanks

[autumn jolt]

This is intriguing. How would you work this into, say, a verse? Like...

[Verse, Softly: 3]

Or two tags on top of each other?

[Verse]
[Softly: 3]

[autumn jolt]

I try to use the double combo of [Outro, Fade out, Finish, End] tags in the lyrics field and also those keywords in the style prompt. I find Suno goes through waves of giving me 1-2 min extensions but then a run of 10-20 seconds endings.

[hushed island]

hi 🙂 how can i reach the support ?

[bronze helm]

On v3.5 (tried it with jazz earlier today) what worked was writing:

[Chorus]
*lyrics*
[Softly] *final line*

But adding numbers actually made it less likely to get soft vocals

[soft gyro]

Hi there, question. After I concatenate 2 songs via Get whole song after expanding there is sudden stop in song at the end, althoud the length is not even 3 minutes: 2:19 minutes. What is wrong?????? 😢 🤔

[bronze helm]

Usually you can write about most things in the existing channels - for bug reports there's #🪲┃bugs-web , for new ideas there's #💡┃suggestions-web and so on

[bronze helm]

Can you check and confirm if the extension ends abruptly? Usually combining parts should produce a seamlessly stitched song

[soft gyro]

the stich is ok, the end finishes too early... I did did for 5 songs: the same! It cuts the song at the end !!!! usesless...

[unique stone]

Are you using "get whole song" on the most recent part?
If so, does that part also end suddenly?

[soft gyro]

yes, I use "get the whole song" on the part2, the most recent one. All end abrubt at around 2 min. I never managed to get the thing right !

[grim elk]

If part 2 end abruptly so does the "get the whole song"

The button just puts the parts together as genererated

[grim elk]

Can you send the song and Part 2? I do assume, that part 2 ends aprubtly and thats why the whole song will end the same

[fluid robin]

male voice, guitar, drum, drum and bass, atmospheric Rockabilly [old] [Funk] Reggae., aggressive - in style of music for a chill song

[charred merlin]

[e-mail] if you want an official response

I'll ping you when I get a response back ❤️

[balmy crown]

Do you think version 3 did this genre any better, or is it likely a coincidence that I’m struggling? (Sorry if this is an odd question, I just figured if it was a version thing, I could roll back to 3 to get something closer to the sound if necessary)

[cold crow]

Does anyone know how to end a song on V3.5? I tried to do it [Fade Out] but it kept going, I tried to do it as instrumental on a point the music sounds like it ended, but it instead it just kept going even still singing even though it's instrumental and lastly I just tried it as custom lyrics and put [End] and still nothing,

[naive quail]

Extend it using v3 from the point that you'd like it to end and include ONLY the [End] tag for the extension. It's still not likely to end immediately, but you should get some options with much shorter endings.

I think my shortest complete track is ~46 seconds using that method

[cold crow]

Thanks that did it

[broken lotus]

post examples.

[broken lotus]

I am DEFINITELY still using v3... switching between v3 and 3.5

(not so much v2 haha, but it has it's moments when I just want a loud electro beat)

[winged sierra]

Hey, Im having a problem i cant seem to find an answer to, I keep getting the "An Error accured while generating a clip! Please try again"

Or it justs says "An error accured generating a clip, an error accured", anyone know of any down times or similar?

[loud nacelle]

Does anyone notice that creating the song with v3.5, extending it with v3, and combining it cleans up some of the "noise" in the song, and the "Full Song" shows as v3?

[gusty acorn]

Hi, I've tried to get some verse to be A Capella with [A Capella] or [Vocals]. But it's not working.

Is there a way to make it do only vocals for a verse?

[broken lotus]

that 'noise' can happen in v3 too but the clips are shorter so they aren't as evident... always Extend from before that noise sound.

[loud nacelle]

OK, so it does clean it up some. Songs sound cleaner when I do the mix but wasn't sure if it was a known....feature, bug or whatever.

[broken lotus]

yeah...
It's a FEATUREBUG 🙌 🐞

lol 😅

[clear escarp]

i really wish the site was better at recommending me things it think i might like, based on what i make and what i like

i think it'd help with discoverability in general

a "for you" section would be massively appreciated 🥵

[clear escarp]

also, the album art doesn't seem to update when you go to download a video, but i'm not sure if this is currently known

[broken lotus]

yeah..., our art isn't actually used in the video... ☹️

[clear escarp]

being able to give it a couple pics that it decides to switch between would be amazing too imo

for the video itself

if it had a suno logo and url and handle on it it'd be amazing marketing as well

as well as being able to download landscape, square, and 16:9 for various platforms 💦

[jade mist]

Can’t wait for V4. Current suno voices sound too AI

[broken lotus]

new feature (in alpha testing now) will allow us to upload our own voices... 😎

the ai voice is based off the 5sec upload.

[glacial lion]

And what is done with the uploaded voice data? I'm curious from a privacy perspective

[broken lotus]

it lives in your Suno acct as a Suno clip... so there are some privacy concerns they will need to work out.

[naive quail]

I'd be hesitant to use my own voice, at least until the privacy details get ironed out. My guitars, tho... they're getting suno'd

cat, too! lol

[pure bay]

Does the metatag [Narration] mean monologue?

[lilac garden]

Will the talked about in painting feauture coming up in the future allow to fix mispronounced words?

[warped furnace]

Dont' know if I want to listen to my voice singing. 😅 But the feature sounds interesting.

[broken lotus]

'Narrator' is the person... if that makes more sense to use.

[broken lotus]

I haven't seen this one yet...,

the way I assume it will work is we take a section of the song, and select some parts that are 'locked', that we want to keep....

the clip will re-render and everything left un-locked will be overwritten (but listening to the locked parts as reference)...

but with the option to prompt the new lyrics...

that's based on the Dev's calling it 'in-painting' like imagegen.

(otherwise it's an erase tool, but same idea, we just erase the bad parts and regenerate

[pure bay]

https://www.suno.wiki/faq/metatags/voice-tags/
It's [Narration], not [Narrator]

[tacit thorn]

Is there any development planned where you could input extant musical samples and have it be a part of the song generation?

[broken lotus]

yes, just posted above
#📚┃suno-school message

[tacit thorn]

Would this also work for musical sampling

Composing a theme or something

[broken lotus]

yes. it works like a Suno clip, we can extend from it.

[tacit thorn]

Ok, so I could use a synthesizer or midi and create a musical clip and then use that to influence the piece

Really nice

Thanks WC

By extension, could this lead to being able to use a full song to create a like-grouping of stylistic songs, like an album?

[broken lotus]

When you start to use it..., yeah, it will take some brainstorming what is possible and what can get the best out of it...

like a really simple idea is to set a tempo with a 4-beat clicktrack..., or the opening bars with the chord progression... signature synth sounds... drums from garageband.... exotic instruments, weird ambience or environment....

Sooooo many possibilities, I think everyone will come up with their own technique...

[tacit thorn]

Yeah sounds really cool and will help with musical direction for folks

Thanks for the help!

[clear escarp]

more discoverability on the site would be amazing, i feel like i only get people to listen via discord even though they often end up liking my music 😩

the newest section fits like 3 songs, and the rest is for like the top 5 most popular songs ever on the site with 10 million plays each

loving the site tho, weirdly addictive lmao....

[gusty acorn]

No-one? Not possible?

[clear escarp]

i guess i don't use the explore much, but it seems full featured... i'm not sure why i don't

[broken lotus]

I have never tried..., so I have no idea if it will work or be reliable at all....

I would try
`[pregnant pause]

[a cappella]
...`

pregnant pause has worked to stop the music a beat... the tricky part will be to get it to go into the a cappella ...

maybe check the hints here about 'sudden style change',
https://www.suno.wiki/faq/making-music/extend-from-time/
if you can get the pause, try extending from that time, and spamming the style
a capella, a capella, a capella, a capella, a capella, a capella...

[lilac garden]

Thanks!!! Im thinking more in terms of Mijourney were u can regenerate certain regions, so more or less that idea. But this would fix a lot of songs i really like but have some mispronunciations. Redoing them right mow through early extends redos the rythm and doesnt give u the same song.

[charred merlin]

I'd try v3 (since it seems to produce a wider array of styles, where 3.5 might lean into making a fully backed song), generate a handful of "seed" songs, and extend the ones with the most promising intros before the music kicks in. Either extend with 3 (if you're still trying to change the style) or 3.5 (if you've "secured a style" and want to expand on it)

Trying to get a complete generation at "the pull of the Suno slot machine lever" can be tricky if you have a very specific sound you're trying to create, as it doesn't matter how specific your prompts are, there's still a degree of ambiguity in the interpretation, but if you make incremental extensions (making sure to get the Whole Song to continue your extensions from so they match the feel of the entire song, not just the most recent "clip"), you can create some wildly specific stuff

[gleaming mica]

does bark have a selection of voices to choose from or I have to clone voice to get any voice to use?

[final hazel]

Bark is opensource, and no longer being worked on (by Suno). This discord is mainly focused on text to music (https://suno.com)

[gleaming mica]

is there no technical help for bark on this server?

[blissful thistle]

First, I really LOVE this product!!!! Quick question. Any idea why the sample rate for music seems to be less toward the ending than in the beginning? Listening all the way through is not always easy to detect, but skipping around makes it noticable. Hopefully this can be addressed soon. Thanks!!!!

[blissful thistle]

Just to add... along with the sound quality degradation, the overall background music in a song seems to go mono and become less present in the song (along with overall quality).

[charred merlin]

Make sure you click ... > report > bug > low audio quality and downvote any songs that have notable instances of audio degradation ❤️ This helps make Suno better for all of us!

[gleaming mica]

can bark be used in monetized youtube videos and games and anything else monetized?

[broken lotus]

yes, Bark is opensource..., there may be some requirements to give credit, I am not sure.

[wintry saddle]

https://suno.com/song/6639495c-6610-4111-8bb8-01db92ca1434/ <-- I'm quite sure suno 'can' follow emotional tags - but like V2/V3 was weak at meta in general, it seems still to be blooming in its ability in v3.5 - not sure what the technical design on this is? and obviously with many spins you can 100% get what you want (the example I asked for far too much just to see... but it does do a lot... )

just seeking confirmation if emotional tags 'do' work - or if I'm just hallucinating 😄

[broken lotus]

which is funnier... [sulk] [sulk]
or how proud of itself Suno seems at the end, LOL

[lapis edge]

Can someone tell me does the ai make the song length on its own? I generated two songs, one is 2:51 minutes long while one is 4:00 minutes, so I just wanna know

[wintry saddle]

Its Annoyingly close at Clapping on demand too

[wind mica]

was there an update to Suno last night where the copy link doesn't link the video anymore?

[lapis edge]

Anyone? It sometimes cuts abruptly when it generates till 4 minutes, which is why I wanna stop it before

[wind mica]

It cuts at 4mins because that's the current max length song it can produce. You can just extend and finish the song.

There is no way to limit the song length

[lapis edge]

Sure but I would like if it could like end before it

[lapis edge]

But one song it made is 2:51 minutes long, is the length randomly decided by the ai too?

[wind mica]

that's what it seems like to me but it's also dictated by the lyrics and song structure in custom mode... if you have enough lyrics or instrumentals, etc. you can get it more consistent in length.

[lapis edge]

I see, so it's basically random right? I suppose the genre stuff I put in also affects it?

And it takes credits to extend the song too right?

[naive quail]

More than one way to skin that cat, but it still feels like a bit of a crapshoot. https://suno.com/song/9cfc0386-c95d-469e-8c90-565ecc2b92b1

[wintry saddle]

@naive quail yeah it was confirmed it derives context from phrases too... (If my memory serves me!) Does. A good job with 1st one... The angry is a total miss here.

[wind mica]

Yes

[void obsidian]

i have been testing 3.5 i love i can place a whole song in now but is there any word on fixing the vocals. i have tried 3 different styles alt rock, americana, and pop rock. each of these the vocals are very robotic.

[gleaming mica]

who I have to give credit to?

I am mainly using the voices it comes with rather than cloning

[broken lotus]

It depends on Bark’s license. Common to need to credit opensource

[gleaming mica]

On the github page it says MIT License but on hugging face somehow it is a different license

confusing

[broken lotus]

🤔

[gleaming mica]

this is on hugging face

it is CC-BY 4.0 NC on hugging face

but MIT license on github

which one is correct?

[broken lotus]

Where did you download it?

[gleaming mica]

github

but on hugging face it is faster to use

[broken lotus]

That’d be the one, i guess

[gleaming mica]

so I can use it for commerical use without crediting anyone for monetized stuff?

[drowsy lotus]

guys, in a pinch here, i've been trying to subscribe but keep failing, (the card issuer treat suno as high risk i think?) is there any other way to subscribe? w/o credit card? will using vpn work? i heard some country can use other methods? been googling but found nuthing 😭

[elfin goblet]

Hi, I'm trying to make a full rap song using version 3.5 of the model, but after having spent a bunch of time and credits, I cannot achieve what I want, I still have sung part! I've tried to fill up the style field with "rap", "hip-hop" and the like, but no way. I've input my own lyrics, and correctly (I think) structured the text with tags like [chorus], [verse] or [bridge]. I know the generation can be unpredictable, but any ideas to force rap style all along the song? Thanks. the text is in French if it can be of any help. 🙂

[final hazel]

ATM you can only use credit cards or debit cards. There is an exception for USA and Germany. There you can use google wallet. More paying options are coming soon.

[drowsy lotus]

oh well, i guess i'll wait patiently, hope it wont take too long, anyway thankyou for replying @final hazel

[drowsy lotus]

try [verse rap] ?

[final hazel]

You can think about getting a debit card. Those are often free and save since you only put on the money you want to spend.

[drowsy lotus]

mine got rejected, called the issuer and they said our system deem the transaction as "high risk" and they refuse to budge -_- smh

[final hazel]

That's unfortunate. I believe Suno uses Stripe, so it's more a problem with them then Suno tbh

[grim elk]

Has anyone tried making throat clearing sounds in suno? And if yes how were you able to make it?

[naive quail]

Not very well... and not very repeatably. 😦 https://suno.com/song/b27a656e-63d3-45d3-a644-795fe9e71fad

[grim elk]

Hm :/

[spare stump]

I see many just use Gansta Rap for a tag

https://suno.com/song/a662898e-e94f-4678-b35c-7bd05c6302d2

I also used Gansta Rap, Drill Rap, Deep Bass ..i used this on my Greek songs

[vapid forge]

Re: Speech output with Bark (not about Chirp)

Q: Could Bark implement interactive voice like GTP-4o?

I'm looking for Open Source (commercially usable) tools to emulate the low-latency speech interaction of the new GPT-4o, which can laugh or giggle, and sing when asked to. Bark is the only Open Source solution I've found (MIT license) that can add a laugh to the audio. I would be hooking this up to an Open Source LLM, like Mistral, Yi, etc.

Can Bark sing? Is Bark low-latency? Has anyone tried this? Any info is greatly appreciated, thanks!

[vapid forge]

From the GitHub README: "Bark is now licensed under the MIT License, meaning it's now available for commercial use!"

IANAL but that seems legit to me... I want to use it commercially as well, if it works for me

[normal patio]

If extending a song multiple times, has anyone had better luck combining the previous parts together before extending?\

[vestal rapids]

use "Get Whole Song"

[normal patio]

Yes, that's what I'm asking. If that seems to make a noticeable difference

[vestal rapids]

it helps alot

[elfin goblet]

Thanks. I'll check it. I've tried gangsta rap in the style field but not as a tag. Should i use brackets?

[unique stone]

This is generally a good thing to do. It's previously helped with avoiding quality going bad and I think it also means the AI is able to see some of the earlier context.

[elfin goblet]

Thanks. i've trie verse and rap with brackets separately but not as a «full» tag.

[elfin goblet]

I'm wondering if the language is not causing a problem here? Maybe it works best in Englsh...

[broken lotus]

A Mod here named Jonathan Fly has some Bark pages, I think the link is on github.... I believe he has many examples of it singing and being expressive...

[faint wigeon]

Anyone know the reason why the style field has such a short character limit?

[vapid forge]

The examples link on the GitHub README is broken

[broken lotus]

there's probably a token limit... the character limit is an attempt to approximate that word-length. Probably longer style prompts would just be ignored after the token max is hit (similar to Stable Diffusion)

[faint wigeon]

Oh thanks, yeah that would make sense

[untold briar]

GPT-4o is genuinely multimodal so it can actually hear your audio, and generate native audio, without any step in between. "Hello" (normal voice) will be different "Hello (screaming angry voice) so using any TTS would not be the same thing, it gets converted into text.

That said, there is a sense in which Bark can actually do the same thing. Meaning you could talk to Bark using voice cloning, insert just enough artifical tokens to prompt a "dialog/voice switch" and then not prompt Bark, but let Bark the audio model generate both the audio and the content (no text prompt, remember). This would genuinely more similar to GPT-4o in that sense. I wrote down weeks ago "recreate GPT-4o demos, but with Bark" a few weeks ago and would love to find time to do it. It won't work in the sense of giving you useful responses, but it would work in the sense of understanding angry versus not angry, and so on - and that would be sort of novel at doing that. Genuinely different than talking to an AI bot through a middle step that convert text-to-audio and autio-to-text. Also I'm 100% sure it would be hilarious (and basically talk back to you like this: https://x.com/jonathanfly/status/1650001584485552130

Also I don't know of public forks that have Bark fast enough to be super low latency, anyway

[vapid forge]

Thank you for the info!

I don't need multi modal input, unlabeled text is fine, but I do need low latency... 😦

[untold briar]

You can stream Bark in theory, but it's a pretty messy and haven't seen an implementation that does it. Except the last two models, that part I tested streaming, and it's okay, but you still have a big startup latency.

However if you don't need super latency, using the small coarse model, a fast implementation, and relatively short prompts, it may only be a couple seconds. But quality suffers a lot with short prompts, loses a lot of naturalness.

I haven't checked in on Bark lately, it is not impossible that huggingface bark or bark.cpp has implemented streaming, actually, worth a quick search if you haven't.

Bark doesn't really sing, unless you really mess with it in ways that would not work well for real time either. Like for example you might generate 10s of audio and Bark stops singing after 5s. So you take the 10s of already generated audio, load it back into bark up to 5s, then ban (penalize really, or Bark will stutter) all the tokens in the last 5s of audio where Bark decided to stop singing. Repeat until Bark has no options left but to sing the whole 10s. looool. (It is still probably very out of tune.)

[fluid robin]

[Saxaphone Big Band] [Drums&Snares]

[Saxaphone Big Band]

[Tuba Rock] [Drums&Snares]

[Saxaphone Big Band]

[Trumpets Rock] [Drums&Snares]

[Saxaphone Big Band]

[Drums&Snares] [Drums&Snares]

[Clapping hands] [Drums&Snares]

[Guitars riff] [Trumpets Rock]

[Pianos Riff] [Trumpets Rock] one that works with any problems

[still basalt]

Hi everyone! 👋

First time asking a question here, having used Suno in the past couple of days.

Is Suno's technology at all like image diffusion models, where there's a random seed per generation? With Stable Diffusion for image generation, you can produce the exact same image if you know its seed and you keep the prompt intact, or a similar image if you retain the seed but change the prompt.

Sometimes I get an output from Suno that I'm really happy about, apart from a few things, and I wish I could keep the core (orchestration, rhythm, etc.) but make some other changes/fine-tunes.

Is that possible at the moment? If not, will/can it become possible in the future?

[final hazel]

Suno does not hold a seed. The way the music AI works is different than SD.

But anything is possible in the future.

[still basalt]

Thank you for the swift reply!

One example is being able to just change intonation of certain syllables, while keeping the rest of the song the same.

Stable Diffusion has seen huge improvements from the implementation of relevant papers since it was publically released (such as keeping a character consistent between generations), improvements that back then were wishful thinking. Hopefully, we'll see similar progress in audio generation. 😊

[bronze helm]

I believe the v4 model which is currently in the works will introduce inpainting, which will allow users to regenerate specific parts to fix mistakes.

Also I have a question, is it possible to download songs as .wav files on mobile? I can't see the option to get a .wav file and I only get .mp3s; my PC browser has both options (mp3 & wav).

[broken lotus]

maybe if you switch your browser to desktop view...?
(beware, the wav are 10x bigger)

[autumn perch]

Why it does not stop when I put an [End] on an extension? It still continues generating nnonsense until 2:00

[broken lotus]

3.5 is a diva that won't get off stage.

I switch to v3 for my endings. or trim in post-edit.

[neat nexus]

I ended up switching to v2 to end it out of sheer desperation.

[narrow cypress]

See my messages in #💬┃general-chat

[narrow cypress]

To be experimented

[peak plank]

Anyone else experiencing quality loss when merging v3 with v3.5? As in, old parts which used to be good quality become bad after the merge. Goes in both directions as far as I can tell.

[autumn perch]

Done

https://suno.com/song/a98879af-0eab-495c-a310-e320c08ee41f

Also with V3 you could generate an image putting a description after [End], but in 3.5 the image description is sung!!!! 😄

[torn zenith]

Hey guys! can I use the music I create in suno in my tiktok channel and maybe make money out of them if people like it? I'm making my own lyrics

[unique stone]

Songs created during a paid Suno subscription can be monetized.
Songs created as a free user cannot (and you must also credit Suno if you make the generated songs available anywhere or use them in anything.)

[tulip blaze]

can any one let me know what would be the prompt, style to be choose to generate a sound clip "slow chanting song without music(without beat sound, slight background music will be there)"
Ex song mention below (https://www.youtube.com/watch?v=I8dmIN46q4Y&list=RDI8dmIN46q4Y&start_radio=1)

[drowsy trout]

Hey guys, trying to do a few hardcore/metalcore songs with mixed singing and screaming and I can't seem to figure out how to prompt the change between the two. Checked the wiki and didn't see anything but I could be blind

[torn zenith]

thanks! also I dont want to monitize it at the moment, I just want to do music to help other but maybe in the future. thanks for letting me know! if i have to pay anything I will to be able to monetize the music if it gets popular, also my channel will show suno promotions

[glacial lion]

It can be hard to get it and you might have to try to gen each section over and over. Have you tried [screaming] or [yelling] in the lyric box?

[jagged sand]

hi

is the pro version worth it

[glacial lion]

Depends what you want to do with the songs and how many you want to make

[jagged sand]

i want to make a parody of ai w a ai musical made completely w suno,udio, runway, about a hypnotist deciding if ai is similar to them and if the ai stealing content and them stealing minds are the same

[glacial lion]

That sounds like it would take awhile with a free account as you only have so many credits

[jagged sand]

yeah'

[full lava]

(only need the first 20 seconds of it)

Any idea why its using an old generation of the verse rather than the new?

The line in question is the last one in verse 4: It did it for 10 sets (20 generations)

New: A new era of music, bringing everyone to their knees
Old: A new story in music, let it please
https://suno.com/song/73435f31-adb6-4a62-8163-b5a0fb2ef7e1

[past cliff]

that's hallucianation for ya

[naive quail]

I think I've only had the stars align perfectly for one song. All the others have little hiccups like that to some degree

and it was the last one I did. heh

[bronze helm]

Right, that's something I could do using Chrome. I forgot. Still, has the dev team perhaps shared any information on whether they will include the option in mobile view anytime soon?

[gentle wasp]

Hey I think I lost 1k credits

scam

[hollow hare]

I assume your credits renewed?

The day-to-day (free plan) or month-to-month (subscription) credits refreshes rather than accumulate because of resources allocation restraints (and is mentioned both in the plan description and at the bottom of the subscription plan page). Suno is not unique for this, most AI services follows this model.

Extra top up credits and special credit gifts are, well, special and are carried over unlike the former two

https://media.discordapp.net/attachments/1069381916492562585/1238249445192237106/image0.jpg

[gentle wasp]

I had a month of subscription and thought I had credits left over from then but I may have went trhrough them I just can't recall

Thanks

[slim summit]

how do i stop the thing from generating the middle of the first line vocals at 0:00? [instrumental intro] does nothing

[untold briar]

It should not be happening every time you generate, strange. If it is, maybe it's something about your style prompt - does the style prompt use any of the same words in the first line of the vocals? Or are your lyrics very long and complicated?

Adding anything sung even just Adding "(ahh)" would probably fix it regardless, maybe empty lines with a period . alone.

[polar abyss]

@full lava probably struggling with the additional words throwing off verse cadence so it's hallucinating

[full lava]

I ended up deleting the main and it still did it for the next 30 generations. It finally stopped but damn

[polar abyss]

Hmm I'm not understanding what you deleted

[full lava]

what i linked was a part 2 so i deleted the main and all the part 2s

unless you mean lyric wise. It was just one line that i replaced

[polar abyss]

But yeah, repeated generation is kinda mandatory to get the result you want

[full lava]

im aware but not one in 50 generations fixed the issue

51 and 52 ended up fixing it

[lapis edge]

I am trying to extend a song but it's not working, I do have credits. It just loads for some time and then nothing happens

[rocky blaze]

your updates are not good, sorry any creation i make is giving tones that are way off, not consistent and not making sense. can you please recover this?

[lapis edge]

Anyone?

[rocky blaze]

i can extend, but the tones are way off, i dont know what updates they are doing but this sounds really bad. i have used 200 credits now and all parts are really bad

[pure bay]

Hi, does anyone know, by v3.5, how to end a song with a natural melody (no abrupt stoppage or cutting)?
I tried metatags like [End], [Fade to End] in lyric box, but hardly worked.

[tulip swift]

Is there a library of working [] commands? Also why does it pick and choose what brackets to use

[fair void]

Hey is it possible to get the text to music or not ?

Like the model

[narrow cypress]

just found out that a rough sounding female voice is actually a man, according to Suno

consistently getting male vocals

[left jasper]

outro

[cold zenith]

finish it with V3. [end]

[signal summit]

Does anyone know how to use "Sampling" in Suno?

[pliant idol]

Hey fellas, does anybody know, how to make singing voice cleaner? I know, that ai model is far from perfect, but some of my generations had crispy clean voice, and some just straight up machine babbling. Is there any way to make it cleaner, any specific voice, that is more preferable right now?

[latent saffron]

Pretty sure it's totally random.

[pliant idol]

So, that means, that i just forget about any "presets" right now and just try to catch a clean voice, like some kind of fisherman? Of course, until Suno release better model with upgraded voices, if they do.

[latent saffron]

Is it even an option? I thought it's model property like 24kHz or 48kHz.

[latent saffron]

Or do I confuse it with sample rate?

[latent saffron]

Original suno tts model has many pretrained voices with different properties. I'm sure v3.5 has something like it, too. But i've never seen it in docs, so, who knows? Just in case, original tts has pretrained voices named en_speaker_1 and so on. You might actually find something like this in v3.5 Or just waste your scores for nothing. Both are good.

[ruby trench]

is there any way to generate only one song at a time? (instead of two) I don't want to burn out my credits to fast 😄

[vestal rapids]

if you only want an option for one song maybe make a #💡┃suggestions-web

[ruby trench]

sure, I started with asking about it before spamming at the other channel 😉

[glass field]

Has anyone got a trick for some keywords to get a vibe similar to Tom Jones? I've tried a lot of stuff and just haven't quite got there.

just want that retro, hilariously dated, baritone male voice, r&b dance pop vibe, but it sounds really modern-poppy when i try keywords like that

[unreal bolt]

how long will it be until its possible to prompt for specific chord progressions?

and also to properly adjust instrumentation, cause right now these parts are just omitted from the prompt if the model feels like some other part of the prompt implies it

[glass field]

knowing*(Having some vague idea) how models like these work that sounds like a long shot

[unreal bolt]

for example if i ask for guitar only, but also kpop influences, it will not only not have kpop influences, but also end up including drums

[unreal bolt]

yeah honestly this really feels like the stuff that shows me that despite how progressed it seems at the surface this is still infant technology

its really just super nonspecific inputs and then trial and error until you find something that vaguely fits your description

i rarely get an output that even remotely resembles the input as soon as its a combination of things that are uncommon

i honestly wish there was a way to reverse a prompt

as in hand in a song that i actually made and figure out what prompt wouldve achieved that

[glass field]

we can't apply other tools logic to this tool

you've got it upside down

[vocal igloo]

How to low the beats behind ?
I mean the vocals are low and beats and music are high

[unreal bolt]

wdym?

[autumn jolt]

Well there's the wiki, but it can be hit and miss whether suno follows the instructions or not! https://www.suno.wiki/faq/metatags/song-structure/

I find that most of the time it does give you some control of the song structure, so it's worth doing even if it doesn't work every time.

https://tenor.com/view/anchorman-paul-rudd-works-everytime-gif-10762553

[unreal bolt]

i feel like it should definetly be possible for an ai to give you the input that wouldve been required to approximate a certain output

why should that not be possible?

basically just using an audio example as the prompt rather than words

[coarse stag]

Hey,
how can you prevent the song from repeating at the end? [Outro] and [End] are sometimes ignored and the last verses are repeated. Does anyone have a tip?

[naive quail]

extend it using v3 from the point where you want it to end, or a few seconds before that point and have only [End] in the lyrics. Pick a good short one, and merge it back into the original from the ... menu's "Get Whole Song" option.

It'll still go past, but usually be much shorter. Do a few rolls and you'll probably find something usable. I've had some as short as maybe 5 seconds.

that's how I got this down to 1 minute. 😄 https://suno.com/song/ce07429a-0419-4ed2-9820-c6faa4ae8214

(explicit lyric)

[coarse stag]

Thank you. Yes, I know that variant. I thought there was just a simple solution that would stop it from the start. But thank you.

[naive quail]

Maybe v4 will actually honor [End]

🤞

[wheat rose]

I've been testing writing lyrics and not using structure tags at all and so far results are pretty good

no [verse] [chorus] etc just let the bot figure it out

[latent saffron]

Because ai is not what you think it is. For starters, it does not understand neither genres, nor language in general. Our text input is just some fragile guideline, what it could sound like. Not "should". "Could".

[wheat rose]

its still repeating but the dynamics of the repeat seem to be changing

[inner moth]

Every time I do this, I get credits refunded. Is that just because it’s short?

[vestal rapids]

often, yes

[latent saffron]

Doesn't SUNO use BERT model for text tokenization? Did anyone test standard BERT tokens?

[inner moth]

Would it still work?

[vestal rapids]

yes sure, sometimes when I extend a song I get 5 seconds of audio and a credit refund, I still use it with Get Whole Song

[latent saffron]

[SEP] to separate part of data
[PAD] for empty space
[CLS] at the end of data

I would only get the answer tomorrow

[vestal rapids]

what are you on about?

[latent saffron]

I wonder if bert is still part of bark

[vestal rapids]

Suno doesn't use Bark

[latent saffron]

I mean, I'm out of credits

[vestal rapids]

Bark is an open source project now

[unique stone]

It pretty much is in its infant stages tbh - I think Suno has existed for about a year or so? My understanding is that the prompts just give the AI an idea of the ingredients you'd like to be in your song. Then the AI just tries to blend those ingredients up, boil them in a pot and then serve it up in a soup bowl.

The problem is, you can't make it cook it any other way, and it has to be blended up into a soup. You can't then eat it with a fork or chopsticks, you can't cut it with a knife, and there's no other way to present it other than to dress it with a little parsley (make minor edits in another program.)

Maybe that makes things clearer. Or maybe it just makes you hungry for soup...

Eventually the technology will improve and actually be able to understand more of what's being asked of it.

[latent saffron]

No? Well, tokenizer still looks like it uses bert

[inner moth]

Thanks! Wish I had thought of doing this sooner rather than wasting more credits.

[vestal rapids]

no, it's not

[undone sleet]

Is there like a way to force the ai to use a specific singer? Aside from male or female i mean. During all of my generations i think ive heard most of the voices but theres this one female voice that i really just like and want to keep using her but have no idea how to force it.

Aside ofcourse from repeated generations and just hoping for the best

[latent saffron]

Totally no. And most likely you can't even make it to use the only voice for all song.

[final hazel]

You can do it soft off with the new upcoming feature: https://youtu.be/6rzt1WNGUSc

[latent saffron]

Oh? Will it work in at least 1 times out of 5? I tested this feature with fazebook musician, and it was... Not satisfying.

[final hazel]

This works like a charm to me.. You can find more examples on my youtube playlist here..
https://youtube.com/playlist?list=PLi8taL4TbG_9x9FH7KZvTaQAD65mSq8t2&si=oNNhnzh9FJGtmPZb

The more you feed it, the better it will mimick.

[undone sleet]

At this point ill just go with making several and hoping for the best

[latent saffron]

I don't like showcases as they don't always present the technology as it is. Not blaming nobody. But getting satisfying result in, say, stable diffusion takes from 1 to "screw it, i'm out!" generations for me.

[unreal bolt]

"Understanding" itself is just a concept. You as a human also just know things in reference to the "training data" youve received over your lifetime. The ai does "understand" that certain data corresponds to certain other data, the same way your brain does it.

your brain does not "understand" either by your definition. Its just when it gets a certain input alot of complex chemical reactions and electrical impulses create a certain output.

[final hazel]

You will see soon enough yourself then. 🙂

[glacial scarab]

I did you come to this conclusion?

[latent saffron]

Wrong. Person knows what guitar is. And person understands what guitar solo is. AI don't. If you want to make AI understand what guitar is, and how to add it into song, you need multimodal system. One ai for guitar, one for solo and one to unite them all together.

[unreal bolt]

You only understand what a guitar solo is after youve been told what a guitar solo is

guitar solo is just random letters until someone gives meaning to it

If i go into the amazon and tell someone that has never been in modern society "guitar solo" that mf dont know shit about what that is

[unique stone]

Suno basically will only remember some amount of audio that extend from, so this is a bit of a hacky workaround that means you'll have to do some external editing:

Extend from the some point during the singing, and then prompt for what you want next. Try to change genre/instruments etc. as you want. It's difficult and the AI might try to stick to the original song, so put [instant switch] or a [bridge] in the lyrics box to try to push it off-track.

No guarantees about whether this will work. The voice might still change. I've been trying to convince Suno to use the same voice but tweak it a bit, but it keeps reusing the same one when I do this, and on another occasion it suddenly switched female to male vocal.

The upcoming audio upload feature 3Daizy mentioned above should help with this though - then you could then take a snippet of the vocal you like, and reupload it. The advantage of that is Suno won't "hear" the rest of the music before/after it so it should be easier to take it in a new direction.

(Sorry for wall of text, I think the coffee has just taken effect...)

[unreal bolt]

cause he cant reference ANY data

you can only understand if your brain is able to reference data its been fed

and for the second part of your answer this is also not necessary whatsoever

in the end music is also just 2 waves for 2 ears.

there is no multitrack seperation in the resulting output

[latent saffron]

1'st, suno's bark used bert to generate tokens. 2'nd, there's no reason not to use it as it is good. 3'rd, I added [SEP] tokens into lyrics, and in many cases there were pause between words in song. Like. Each. Word. Were Sentence.

[unreal bolt]

your brain is simply able to discern what is each element based on having leart it prior

[latent saffron]

So, even if it's not bert, it uses alike token syntax.

[unreal bolt]

and if you dont know an element then your brain might confuse it as something else aswell and/or hear it as a combination of things

for example im musically trained so i can hear what notes make up a chord in a song, but if youre not musically trained you experience a chord as a single entity sound.

and whatever you feel about this sound is also just learnt behaviour. there is nothing inherent about it that cant be quantized.

a minor chord sounds sad to you cause youve heard in context of things that you were told are sad.

and this is essentially the exact same thing a neural network does: you tell it "sad" and then it references its data that has been marked as "sad" finds a minor chord progression and voila output

[latent saffron]

Well, even if one don't know the word, he most certainly will understand the structure of the song. So, you only need to say "this part is song, that part is solo" to make it clear. Our brain does deduction and induction of all incoming data.
Current ML models don't. Neuron in ML is just too simple compared to human's brain ones.

[unreal bolt]

but thats also learnt behaviour

[unreal bolt]

yeah obviously i agree that its a complexity problem, but its not functionally different

and ye llm dont constantly retrain on their input data

they are basically a snapshot of a brainstate

[undone sleet]

Oof welp. Guess my new upcoming album eternal embrace will be having different singers then

[latent saffron]

it's not about learning. It's about available functions. There's programming language named Brainfuck. Which is good enough to write any program. But it's far too simple to actually do so. That's what ML neuron today compared to human neuron. And that's why we are dozens of steps away from actually good text-to-something models.

[undone sleet]

Im sorry but WHAT?!?!
who in everything that is holy would name a coding language Brainfuck?!?!?!

[unreal bolt]

but wasnt that what i said to begin with? what you said basically just means it needs to be way bigger. not that it needs to be functionally different.

[latent saffron]

Well, it's name fully describes it. No one in sane mind would use it . For the exact reason the name states.

[latent saffron]

As we can't make one good big ai, we need to make a lot of small ones. And then combine them in a smart way into bigger multimodal ai.

[unreal bolt]

okay, but this has nothing todo with my initial point and/or you saying that is a functional misunderstanding of the way ai works.

cause it isnt.

it completely possible for a llm to take input data in other forms than words to create its output or even to approximate the word input that wouldve caused it to give an output similar to the material you gave it for reference.

[latent saffron]

Because current AI does not understand what the guitar is. For him, there's some cloud of samples what guitar could sound like. Which includes guitars. But also drums and flutes and human voices.

[latent saffron]

Well, that's what "multimodal" stands for

[unreal bolt]

but thats just inaccuracy. thats actually true for your brain aswell to a certain extent.

if a voice sounds similar to a synthesizer or instrument in a song for some reason you will not think that thats a voice

drums are not part of that cloud cause the AI doesnt "understand" the difference between drums and guitar, but just cause it doesnt understand ENOUGH to always keep them apart.

a multimodal model would fix none of that esspecially if its about reverse engineering an output.

if its inaccurate the drum model in the complex would mistake a guitar only solo as including drums and fuck up the output.

[latent saffron]

We're discussing generative models, no? You can imagine the sound of guitar, no? If not, it's okay, some people can't imagine picture or sound by their will. But if you can, you most likely can imagine two guitars playing simultaneously. And you can describe it "two guitars". Well, AI can't. Because it's neurtal model is too simple. And in current terms you need to add too many neurons for this AI to be used by anyone except IT giants.

[spare stump]

looking good in those videos

[final hazel]

Thanks 😄

[latent saffron]

So, in theory, yes, we can write any program using Brainfuck. And we can teach AI to understand everything and do anything.
But in practice?

[unreal bolt]

I think we are talking past eachother a bit.

[latent saffron]

Nope. You started with "If it's possible to make AI to understand chords?". The answer is both yes and no. Yes, because it's quite simple to take cords, generate baseline sound sample and then use it as song's main line. No, because you need damn good tokenizer to understand not just cords, but which part of request uses chords, where, how exactly. You also need better mixer to mix not only main line, but also sublines and and singing. You also need better control over timeline, so, you need larger memory layer. And so on and on and on. If it were any simplier, it would be already done.

[unreal bolt]

Youre first answer was in regards to why it wouldnt be possible to reverse engineer an approximation of the prompt that created an output.

[unreal bolt]

And you simply said:

[worldly mauve]

I don't get the problem honestly

it's just a matter of training data

[unreal bolt]

thats what i be sayin

[worldly mauve]

if you want the model to understand input X you need to train a conditioning and all you need is enough training data

[latent saffron]

Can you say which part of which plant or animal were used in your dish by looking at it?

[worldly mauve]

technically it's simple

[latent saffron]

Well, no. It does not work this way.

[unreal bolt]

not just by looking but yes by taste aswell

[worldly mauve]

and why?

I don't know what suno is internally. I would bet it's a diffusion model, but there is not much information out there about that. But shouldn't be so important anyways

[unreal bolt]

and if i have actually eaten a similar meal before that im able to cook then i can even approximate the recipe of that dish

[latent saffron]

Lol. Not even DNA test can do. And you can by just look and taste?

[unreal bolt]

wat

i can look at a piece of meat and eat it and i will know what animal it was an which part i ate

thats not even a crazy take lmfao

and the same goes for the rest of the ingredients

do you eat your meal and the entire thing just tastes like THAT MEAL or do the things inside the meal that make it up taste different to you?

[latent saffron]

It is important as diffusion models tend to grow exponentially with every new piece of knowledge.

[worldly mauve]

Write a program that extractd chords automatically out of audio samples. If that's possible you can use that to train the model the other direction 🤷‍♂️
Reverse engineer is the right idea here

[worldly mauve]

never heard about that

you can just add another adaptor layer, e.g. with cross attention. Cubical running time.

[latent saffron]

it is if you get one solid piece of meat. Meatballs?

[unreal bolt]

to a lesser degree still yes

i would definetly still be able to tell you what animal its from unless its made from an animal that ive never eaten a meatball of

thats the entire point here tho

[latent saffron]

Well, maybe exponentially is a bit too much, but go get base StableDiffusion 1.5 model, and then get all loras for it to do anime and toons and NSFW and stuff.

[worldly mauve]

and then...? lol

[unreal bolt]

im able to reverse engineer the meatball as long as i have reference data to a similar meatball ive eaten

[worldly mauve]

loras don't increase runtime

they operate in the same state space as the original model

[latent saffron]

Well, that's the point. You'll need too many resources to say "Yes, the word metal definitely have 90% possibility to make it sound metal". It's near to impossible to reverse it.

[unreal bolt]

what

it has the exact same chance in the opposite direction lol

If the word metal has a 90% chance to create something that sounds like metal, then if i have something that sounds like metal, there is a 90% chance it was created with that word

i mean isnt that essentially how an LLM is trained to begin with? Youre generating outputs based on inputs and then comparing it to expected output data.

[latent saffron]

Than what do i confuse it with? I definitely used something different to get result outside of original model...

[worldly mauve]

you might have used a lora, but what is the problem with that

a lora just retrains/finetunes the model for a new task in a very efficient way

most of the time, though, a lora is not even changing the model's capabilities, just it's prompting behaviour

cause the model is usually already able to do everything - it's just difficult to find the right prompt

but that was the question, right? Can you add new kind of prompts (such as chords) to better guide the model

yes, if you have enough training data

[latent saffron]

Unless you do it outside of it's zone of comfort. Like, generating 2048*2048 with SD1.5.

[worldly mauve]

that's a technical limitation of transformers

and convolution

[unreal bolt]

doesnt a lora just unify inputs to restrict its output?

[latent saffron]

that's exactly my point

[worldly mauve]

Suno is also restricted to 3 or 4 minutes

but that's about length

and it's not "exponentially"

it's quadratic

and adding chords wouldn't be so dramatic

you can dropout information or just use adapters. To stay in the stable diffusion analogy: in SD 3 you can either use CLIP or T5 as Textencoder, or both. So you can choose yourself how much extra information you want to provide

and ELLA is an adapter that allows you to add T5 to SD 1.5 afterwards

controlnets allow you to use images as inputs

yes, all that costs a bit if performance, but its not exponentially

[worldly mauve]

no, it's just a clever fine-tuning technique

[latent saffron]

True, all true. But it's multimodal processor.

[unreal bolt]

yeah but i thought the finetuning was just prespecifieing a bunch of redundant input values so that the output always leaned into a certain direction

[latent saffron]

Go fit it all into two prompt inputs

[unreal bolt]

hmm?

[worldly mauve]

I don't understand what you mean... fine-tuning is training the model on new data

[unreal bolt]

ohhh i get now what your issue with reversing is

you think i expect the EXACT prompt that would create EXACTLY that output

[worldly mauve]

that's not possible by the nature of diffusion

[unreal bolt]

ye

but an approximation would definetly be possible

[worldly mauve]

but you can definitely get probabilities for each word if it "guides" generation into the right direction

[latent saffron]

Nope. You can't reverse weight of any token at all

[unreal bolt]

not exactly but you can approximate it

if your image is the portrait of a woman the chance that the prompt used was "portrait of a women" is higher than "sideprofile of a dog"

and the ai could definetly approximate that

[worldly mauve]

there is CLAP (a audio variant of CLIP) that can do that.

in theory you could also do that on the original suno weights via back propagation. But that's probably extremely slow

[latent saffron]

And you need to start with "@final hazel Sorry for bothering you can you please share your tokenizer dictionary for style input? Thanks, love you!"

[worldly mauve]

in principal that would boil down to brute force trying. Add noise to the sample, check for each word how much noise is removed. Probably too slow in practice. but CLAP should be very fast. It's the same as CLIP interrogation in SD

I guess the problem is just that suno is closed source

so we neither have their tokenizer

nor their text model or anything 🤷‍♂️

[unreal bolt]

Why would that be necessary tho? Maybe what im describing isnt really a neural network, but ai is already able to split music into stems right? So the functionality to tell a guitar apart from a drumkit or a vocal exists. Now these models "know" wether there is guitar in your song or not or rather "how much" based on the output they give and that alone would already constitute part of a prompt that couldve created that music.

If the ai tries to split the guitar from the rest of the song but the output data is almost silent cause there is nothing that resembles a guitar then the ai would know that the chance this piece of music was made by asking for something with a guitar would be low and that there is a good chance it could be achieved by specifically prompting to not include a guitar.

[latent saffron]

Well, SD even with all ControlNets and Loras and etc etc still do not understand neither text nor human body. It does understand what parts hand has, which tokens it activates. But nothing more. No number of fine-tuning, no number of datasets can make it understand the hand outside of part of image. That's why it is impossible to make SD to draw normal hands with normal number of fingers holding something in physically realistic way

Physics is simply outside of it's scope.

[unreal bolt]

yeah but thats because it doesnt have enough reference data

thats the only problem

[latent saffron]

The problem is you need 4 (or more) MLs to do so:

2. split guitar
3. estimate guitar
1. generate music again
4. add generated music```

[unreal bolt]

the only reason why you can tell that something looks unrealistic is cause you have a boatload of training data from interacting with objects in the physical world

[unreal bolt]

i mean im not exactly sure why they sould need to be seperate or what you specifically mean by that

[worldly mauve]

because it's not trained on that. If you would add propiate training data it would learn it. But we simply don't have billion of training data for that.

[unreal bolt]

cause my brain can do all these processes by iterating through the same functions a neural network woudl

[worldly mauve]

human brains and artificial neural networks have basically nothing in common

the whole term "neural network" is highly missleading, as the way it works has nothing to do with real neurons (there are not even neurons in neural networks lol)

[latent saffron]

nope, because you need more complex model

[worldly mauve]

some people think that on a very high level, information processing might follow similar rules between humans and computers, but this is just a theory without any proof or even any hint

[worldly mauve]

not necessarily

llama3 and sigma phi are nice examples that show how good a tiny model can become with just better training data

in SD with control nets you often get correct poses and even hands. It IS a matter of training data

[unreal bolt]

the only difference between neurons in my brain and the transistors in a computer is that the ones in my brain can have fluid states

[latent saffron]

The models we have now are too stupid. That's also the reason why it's impossible to make animations with SD. We need far more complex models/

[worldly mauve]

of course for giving the model an universal understanding of the world it would need to be much bigger

[worldly mauve]

we don't even know how neurons work. But I'm pretty sure they are not like transistors 😉

[unreal bolt]

we dont FULLY know how they work, but from what we know they do share functional similarities.

[worldly mauve]

as said I don't know the internal of Suno as it's a closed model. But I don't think there is something like "this song contains a guitar". I would say such information like "using a guitar" is just an conditioning for the diffusion process. Which means: the model don't need this kind of guidance. it can also create a guitar song without anybody telling it to use guitars. It's just that using "guitar" in the prompt will guide the model to make musik that is more similar to music which was flagged with "guitar" in the prompt

[worldly mauve]

yes, but a chess computer is then also a synthetic brain with that logic

[unreal bolt]

yeah absolutely. i completely understand that just cause an output contains something doesnt mean the word was used to create the output

[worldly mauve]

I don't see much similarities between artificial neural networks and biological neural networks. It's more or less a "marketing thing". People in science do that as well as enterprises, use fancy names to get more funding

[latent saffron]

Well, we know pretty much. Each neuron is at least as complicated as standard Arduino. And can also move. I'm pretty sure tensors for now are a bit too stupid to emulate brains.

[sturdy fox]

I was sharing links to X and got a message on the post saying X refused to connect to the link: https://suno.com/song/b136ae3c-5032-4a81-9e06-a3d5d2c0e51c

[unreal bolt]

yes in a way it is, but its a really simple one that incapable of doing anything but finding the correct move and occasionally purposely make mistakes in order to match a certain player skill

[worldly mauve]

I don't wanna claim that we cannot simulate a brain via computers. I just say that neural networks (basically tensor multiplications) have nothing to do with biological neurons

and arguing with "what a human brain can and cannot do" is therefore always very dangerous, cause it quickly leads to wrong conclusions

[unreal bolt]

imo a biological neuron could be emulated by a cluster of transistors

but obviously the analog nature could never be 100% replicated digitally

[latent saffron]

Yes, the "guitar" subspace. Which also has flute and bass and a pinch of drums.

[unreal bolt]

cause digital states are limited and analog ones are not

[unreal bolt]

yes but again this is just inaccurate training data

if enough people told you that drums and guitar are actually the same thing when you were young youd believe both are guitar aswell

there is nothing in your human brain that tells you that guitar and drums are something inherently different

[eternal plaza]

Is there to download and run the music model locally?

[latent saffron]

Well, yes, but no. Your point is "we can extract exactly guitar". My point is "we can not step outside of that exactly guitar". Having clear "guitar" does not allow us to use it as a primitive to build songs. Having clear "guitar" does not transform into "three guitars".

[unreal bolt]

my point is "we can approximate guitar very accurately" and three guitars is more similar to three guitars than it is to a single guitar

[worldly mauve]

does anyone has a good style list and prompting guide? You find many webpages out there, but none feels comprehensive. I'm a bit surprised that there is no wiki or something in here that contains all tags and styles that worked for someone as well as prompting hints

[unreal bolt]

your human brain also doesnt know why the guitar sounds a bit different than a single guitar until i tell you "thats actually cause its more than one guitar" from that point on you have a reference to what sounds like 1 and what sounds like multiple guitars.

[latent saffron]

Again, you need to start with "@3Daizy🌼 (suno.com/@3daizy) Sorry for bothering you can you please share your tokenizer dictionary for style input? Thanks, love you!"

[unreal bolt]

and the same can be applied to neural network

[latent saffron]

@final hazel also, please, help this poor soul #📚┃suno-school message

[unreal bolt]

as long as i give the network reference data for what is a solo guitar, whats a duet, whats an entire ensemble of guitars, it can tell them apart

[final hazel]

I see.. but that is out of my reach tbh. I am a moderator, not a developer 😅

Not even sure if the problem lies with Suno or X.. but I will send it to a Dev

[latent saffron]

You're one step closer to developers then we are. Also, love you!

[worldly mauve]

@3Daizy🌼 Sorry for bothering you can you please share your tokenizer dictionary for style input? Thanks, love you!

@final hazel Sorry for bothering you can you please share your tokenizer dictionary for style input? Thanks, love you!

[final hazel]

"tokenizer dictionary"?

[unreal bolt]

id die laughing if a developer came in here and just said "yeah this is all very possible and achievable within a few years"

[latent saffron]

It is possible, but not in a simple way

[unreal bolt]

its possible by litterally just scaling up what we currently have

you need no new technology, just more data with more metadata

[latent saffron]

It's list of words style names that are surely known to ML and have their personal ids. Devs will understand.

[final hazel]

https://cdn.discordapp.com/attachments/757388174854193202/1247136004251320390/linasanad-lina_1.gif

[latent saffron]

If impossible, well, at least we tried.

[final hazel]

Now I understand.. I have none.. I just TRY everything @worldly mauve

You can see all my publish songs on what I use.

The model has a bit of everything, so I usually manage to get what I want.

[worldly mauve]

not very helpful 😅

[final hazel]

I just use my common sense, and websites like "https://rateyourmusic.com/" for genres.

My songs are helpful, I think?

[glacial lion]

There are also a bunch of styles here: https://suno.com/explore

[worldly mauve]

can you be huffy? xD

[latent saffron]

Well, known issue with MLs:
ML: * Gives perfect result *
User: Wow!
Dev: WHY??? HOW???

[final hazel]

Me huffy? Noooo lol

I am not easily angry 😉

[unreal bolt]

i mean its just trial an error

most of sunos outputs are complete trash

i dont think its all that unexpected that one in a few thousand outputs will fit the expectations

[dusky quartz]

Where can I get points back for failing to do a simple task like 20x times in a row?

[inland jacinth]

Is there a way to change some lyrics but not generate a whole new song, but use my existing as source?

[neat nexus]

Not at this time. The only current way is to resume from time and change the part but it overwrites the rest from that point on.

[chrome lion]

Guys, any idea for a prompt to make a proper intro into a song?

[glacial lion]

One trick is to make an instrumental first, and then extend after the intro and put lyrics in

That is how I got this intro, and then the lyrics were inspired by the melody: https://suno.com/song/2a638df5-8ead-4855-9ad7-ea97065e4e98

[chrome lion]

Thanks. 3.5 seems to behave a bit different than v3. Never had this issue before.

[glacial lion]

Yes, I was struggling with that same issue when I came up with this workflow. 3.5 seems to jump into it too fast sometimes

[gleaming mica]

in hugging face it was a different license that was why i was asking

also everytime i ask it it uses a different voice

I want one voice instead of a new voice everytime

what is the command for one voice instead of a new one everytime?

[naive quail]

Niiice! That's a good trick! I used that with a freaky instrumental piece to make this one: https://suno.com/song/85ee4c32-47cf-474f-8f65-0e9ad6ea0818

[vapid forge]

Maybe they updated it recently but it says MIT now on huggingface

[gleaming mica]

in the space it was CC

[supple marsh]

Just wanna say this here cause i kinda got little bit of shock, yesterday i logged in and it was almost like i had access to every song as in as my own but logged out and logged back in again and everything was normal

[vapid forge]

Interesting. I think it was originally CC NC but then they switched to MIT, based on the GitHub README. They probably just forgot to update the space.

[supple marsh]

every song on suno that is

[gleaming mica]

Probably. I have a new voice everytime with bark what is the command for one voice?

[vapid forge]

This guy on Twitter says using a quantized 4bit bark-small model he can get audio tokens in "7 ms".

https://x.com/el_PA_B/status/1782372867872547063

Pierre-Antoine Bannier (@el_PA_B) on X

Bark.cpp generates audio tokens in just 7ms! ⚡

Initially, the 4-bit quantized Bark model produced tokens in 25ms.

Switching to @suno_ai_'s small Bark model reduced inference times to 7ms while maintaining great audio quality!

We support both models now!

Twitter · Apr 22, 2024, 11:36 AM

[prime sonnet]

I can not extend the song ERROR "An error occurred.
A system error occurred." What is happening??

[vapid forge]

But he does not explain... Time to first token? 7ms Per token? Not sure if it qualifies as real time...

[untold briar]

It's 7ms for the first token, of the first model, so it's not 7ms until you hear audio. It's also not streaming. However that's running on CPU so it's impressive, and a similarly optimized GPU version could be a lot faster some day

[idle jacinth]

Uhm, to get a guitar solo i simply write (guitar solo) right?

or should it be [guitar solo]?

[glacial lion]

this

[idle jacinth]

thanks

[vapid forge]

The latest Bark GitHub README update claims "10x faster on GPU with less than 4 GB VRAM" so maybe I'll try on Runpod.

There is a video of it running pretty quickly on an iPhone

[soft gyro]

can anyone share the lyrics with the meta tags, of a very complex built song, I mean with a lot o instructions for the AI? Or is there any place from Suno we found that?

[grim elk]

Not sure if the channel name was meant for these questions but: A Friend of mine would like to buy some credits (or get the sub). People in Germany are supposed toi be able to use other payment methods other than Creditcard, but for him its not shown ... does anyone know why?

[soft gyro]

Hi Jonathan, how many CPUs has the AI all together? Is it scaleble ?

[untold briar]

I'm not sure what was used in that video, it may actually have been a macbook using the CPU + neural engine, since that fork is specialized in apple support

[soft gyro]

I don't mean any video, I mean the Suno hardware where the AI software is running...which consist of what?

[untold briar]

You mean Suno music? No idea about that, sorry.

[soft gyro]

I thought you are from Suno...arent you?

[untold briar]

Suno Discord mod but don't work for Suno

[soft gyro]

ah, ok. 10x. Where and who can I ask that in the suno discord ?

[untold briar]

I don't think they talk about specific details like that, kind of the secret sauce I think

[soft gyro]

Ah, I see... probably is secret, but did you ever asked them? Maybe is not... anyway the CPU power has to be increased with the number of user...so maybe they would talk about what they had 6 months ago...

[primal quail]

Hi, why are the vocals sounding so Vocaloid and yoddling since the update to 3.5?

[sturdy flax]

It would be nice if this technical discussion room was more focused huh? lol

[neat nexus]

No real information exists on the reason but the most common speculation is it was a bit over trained?

[weary wharf]

Any idea why >15k harmonics are missing?
Is this happening because of the training uses compressed audio data?

I know most people cannot hear above 15, and general is below 15 and highest around 12 for adults. Just wanted to know.

[neat nexus]

🙋 Possibly dumb question on my part: If you can't hear it any way, what does it matter?

[unique stone]

Is this for WAV or MP3?

[weary wharf]

Normally when compressing, that part of the audio is moved to a lower freq, so it can become hearable. Some instruments can go high, and moving those harmonics improves the quality. If training data uses compressed, it means it's moved. But if we use original uncompressed and this is cut because of what suno generates, than those high freq noises are missing.

Like strings can sound a little bit dull when they go higher noises.

[weary wharf]

This is the wav output.

[neat nexus]

Hmmm...

[unique stone]

Huh weird. When did you download this and which model was it generated with?

[neat nexus]

That could explain a few things.

[weary wharf]

Downloaded a few hours ago. It's using 3.5

[unique stone]

I have noticed that some songs seem to have an overall better quality from the start, and others are poorer. I have some old-timey mono sounding song for example. It doesn't seem possible to break out of that quality, I've found - once the song is that quality, that's it for the rest of the song (which is why v3's annoying metallic noise needed to be avoided before it developed in each song)

[weary wharf]

I think that depends on the starting "randomness" of the audio.

Like for image generation, you generate a "noisy" image, and model tries to fill and convert it to an image/photograph/drawing/etc. Depending on the starting noise, it's either a good one, or a bad one -- not talking about taste, but the content.

Or think ChatGPT. The more you talk, the more it mumbles, looses context, generates more details, unnecessary text, and at some point all goes to 💩 .
It's probably same with audio generation as well, the more you move into the depth, the more "noise" you'll get.

Even if it starts with a really crisp sound, in a few minutes worth of audio, you can start hearing the noise and breaks in the instrument sounds.

It's not about >15k being cut 😄

From experience, more instruments = more noisy continuation.
I think AI when it's looking back for context, captures the noise and continues generating it over and over.

Just piano, you can create a 15 minute audio without it becoming noisy.

[gleaming mica]

Does bark have one voice command?

[weary wharf]

That depends on the genre as well. 70s organ sound will sound noisy definetely, as they are probably coming from remastered magnetic recordings.

Now this is strange. In this "piano + guitar + bass" acoustic, that is created by Suno, I can see that we have harmonics up to 20k. Not cut after 15.

Try to extend your style, maybe something like pipe-organ, film-score, modern-cinematic, pipe-organ, pipe-organ, pipe-organ, short, short, melodic, short, melodic, sad

As long as the sound starts to sound "noisy" just do not listen the rest, extend 1 second before that point. You'll get a pretty good output. But it burns a lot of credits 😄

[eternal oar]

https://suno.com/song/83269d9b-36cf-41f7-a979-981a321839f5 I'm so ready for a crop button or a clear way to command an ending- I've spent 20 minutes trying to end it, but my "fadeout" is interpreted as introduce saxophone

[thin fern]

I'm a drummer and would LOVE to create tracks that don't have drums in them already - doesn't seem possible now, is something like this in the works?

[teal raft]

Easiest and quickest way right now would be to download the songs (with drums) and use an outside program to break the instruments apart and simply combine them back without the drums

Won't be perfect, won't be flawless

But can be done right now. Suno works in masters only

[keen lotus]

Either my speakers or ears are broken or Suno sometimes changes the part I'm extended from after clicking on Get whole song. So the part I extend from also gets unusable if the new part was noisy or of poor quality.

[weary wharf]

Change the second... Go 1 second before or after.

[keen lotus]

I mean that my part1 also get noisy resp. poor quality, after I clicked at whole song. So part1 cant no more used for new extends.

[weary wharf]

If you are using 3.5, download the parts as wav, and the original full song as a reference. Mix the parts into a single song youself. The "generate" does some remastering, which increases the gain, and most of the times introduces noise and clicks.

[rapid bluff]

I've noticed with the 2 gens you get, 1 of them is typically better quality then the other, its not as noticeable in V3.5 but V3 it is/was night and day. My theory is each generation gets alotted a certain amount of GPU resources which is then split into 2 to create our songs. I'm thinking 1 of them get a bit more than the other. Again, just speculation after months of use, I may be completely wrong

[untold briar]

Well, if you roll two dice, one of them is typically higher than the other. It is more than likely just that kind of variance, because I don't think you can just spend more GPU resources to make something strictly better. In some models you can do more steps, but this isn't always better even then. I wish you could just spend more - I'd happily spend 100 credits for 10x the quality.

[torn sun]

I need that too.

Ending songs is really frustrating.

[weary wharf]

I think it depends on the latent space, and how the seed moves in the space. So if the movement is close, you'll get a similar song. If the movement is a big gap, you'll get a totally different song.
I do not know how many features the model has -- you can think it like dimensions. A color cube.

If you move around you are moving in colors near you. But still get a different color. Probably the 3.5's latent space is bigger/higher then 3, so you get similar things. 3 has more difference.

[rapid bluff]

interesting way of putting it, maybe that's what it is? 1 is using more "steps" which keeps the prompt consistency and quality of the song intact. Similar to making a photo with any image generator, typically if you run 4 images per seed, 2 are structurally closer to what your asking while the other 2 are similar but not quite what you were asking from it. Granted all 4 of those images use the same amount steps so just seed variance which might be similar to what @weary wharf is saying

[untold briar]

It is possible to spend more or less steps in some models (no idea for Suno), like I just saw this Diffusion audio model pop up today with an interesting 1-step versus multi-step: https://arxiv.org/pdf/2405.18503 but it doesn't make sense to me Suno would make things work like that. Unless people sometimes prefer lower steps, in which case, it's just a matter of being different not worse anyway

[outer trellis]

The wiki says “How do I get my song on new/trending?”: The song must also consist of at least 2 parts.
Was this changed after the introduction of 3.5 or is it still the case?

Should be extended once or not?

[untold briar]

I'm pretty sure that requirement dropped after 3.5 introduced 4 minute songs, though I should probably double check, going from memory

[gleaming mica]

Is it possible for bark to make 1 voice instead of many? i want to use 1 voice only instead of new voices everytime

[rapid bluff]

have you messed with that yet? curious how good the audio quality is. I agree though, in some cases the wild left turn the other version makes sometimes is what I strive for, and even if its a 2 second cool thing it did before the audio deteriorated, I get the full song just to capture the 2 seconds and continue from there

[untold briar]

Yes you can save any Bark random voice and reuse it. In the Suno repo the function is called save_as_prompt(). In other Bark implementations it is likely different. But the process is simply saving the raw Bark output tokens (not just the final audio) that's it, that's the voice.

[untold briar]

The github for this isn't even up yet. There is a new Stable Audio out today though, if you'r elooking for something to mess with.

[gleaming mica]

is it possible with the command line?

i used python -m bark --text "Hello, my name is Suno." --output_filename "example.wav"

[rapid bluff]

sweet ill check it out!

[untold briar]

Looks like the command line added the flag --output_full True but it doesn't actually work. There are other bark forks that will do it though. https://github.com/rsxdalv/tts-generation-webui or https://github.com/JonathanFly/bark. If you know Python it would be a small change to make that work, I believe.

[keen lotus]

I am not good with DAW, but thats the result. Full song: https://suno.com/song/d4cfd295-2d22-4e53-8c65-686c1fce0099 DAW song:

[weary wharf]

It sounds much better, vocals are on top and the bad noise on top is not tha hearable.

Still the small clicks are hearable in the hard consonants, but that's coming from Suno and you need to go deep in to clean it.

[gleaming mica]

I don't know Python i'm trying to learn it. Do these forks use the latest version of bark?

[untold briar]

Bark has later implementations but not really later versions. In other words, the core model is the same. However I would say the HuggingFace Bark is the closest thing to the latest version simple because by being a part of the larger Huggingface AI library 'transformers', Bark gets some new optimizations and also more sampling options for free. However, by being part of a generic AI library, it also makes Bark more confusing to understand (a lot of the Huggingface language seems to be written for text models, which are most of the transformers in the library).

If you want to learn, if that's the main reason for trying it, I think maybe the Suno Bark is actually better for that, by being more simple. But Huggingface is widely used so learning that library would be very useful, but it is a whole level of complexity on top of regular Python.

I could probably add saving voices to the official Bark CLI quick, if I get some downtime I'll do it and send you link. Probably like <10 lines of code realy.

[weary wharf]

There is a problem at the end. Looks like there is a space in the ending.

[keen lotus]

A suggestion from me is to fix this Suno problem. Because as a beginner, it's very difficult or even impossible to get things like this right with the DAW (the transitions and clicks)

[weary wharf]

You can use fade tools for joining parts... It will drop the clicks on the starts / ends.

Here.... I think this was not merged correctly.

[gleaming mica]

thank you! main reason is i am using it for both learning and monetized video. If by huggingface bark you mean the huggingface space of bark I have tried that but wanted to run it locally

[keen lotus]

Yes you are right. You have tool suggestions?

[weary wharf]

I use logic pro for everything, but I'm doing mixing, mastering, etc... It's an expensive tool.

[vestal rapids]

I use Logic Pro and FL Studio to create music videos

[neat nexus]

FL Studio makes videos?

[vestal rapids]

yeah ZGameVisualizer (included with FL Studio) lets you do effect videos with generative content, here's an example of Suno music in a generative music video https://youtu.be/t3aoR26lwiw

[neat nexus]

Ah, it's an audio spectrum animation thing. Like Vizzy.io

[vestal rapids]

and raindrops!

[neat nexus]

That beaded raindrops was cool though.

[vestal rapids]

if you're willing to play around you can do some cool generative effects like that and every single frame of the whole video will be different

[neat nexus]

I suppose. I mean I don't mind such things but I have a preference towards actual music videos, my current output lately aside.

Though those were fun...

[vestal rapids]

this is why I await tools like Sora, I already use Pika in some projects

[neat nexus]

I've given up on Sora. As others have said, where is it? The longer we go without a public release the more it looks like cherrypicking.

[keen lotus]

now it should fit halfway. you have also fade tool suggestions for beginners (cheap tools)?

[remote ridge]

When will suno get the audio upload and you can extend that exisiting song. Udio just released it its fucking great

[vestal rapids]

it's not going to be released this year

[neat nexus]

I use PixVerse most of the time. (Full Disclosure: I am a mod there.)

[neat nexus]

So it literally doesn't matter.

[light pike]

I cannot for the life of me end the song when the lyrics end. It just goes on for the full 4 minutes, no matter what I do

Just repeats lyrics and then the song abruptly ends at 4 minutes where it makes no sense

[weary wharf]

For DAW's? FLStudio can be a good start. One time payment, life time ownership.

Bandlab has an online free version of something like Cakewalk, but it's 💩 , do not know how good it is for audio processing.

[torn sequoia]

What to do if I don't want a given verse to be sung twice? Sometimes it happens that random verses are sung at the end of the song.

[weary wharf]

Suno does what Suno wants... Just retry extending.

[weary wharf]

Try adding a tag like [instrumental][outro][end] one after another.

[torn sequoia]

at the end of a verse or song?

[winged tusk]

hi ... what is the best way to assemble a song with parts and share it here ?

[weary wharf]

This is as clean as it gets. You cannot recover the lost harmonics that Suno is not putting in.

[keen lotus]

Thank you very much

[weary wharf]

It's Spotify ready if you are going to release it.

[keen lotus]

Maybe at YouTube. You want that I write your name as feature?

[weary wharf]

No worries 😄 You can leave me out.

[idle lily]

Hi all. Quick question. Any tips on how I can assist Suno in differentiating between Live as a "live show" vs. live as to "be alive"?

[keen lotus]

It's no problem for me if you want it

[weary wharf]

It's your decision put if you want. 👍

[keen lotus]

In that case should I use The Builder or another name?

[modern belfry]

Is there anyway to make the AI speak more clearly? I type [outro spoken] It will talk it out, but its not very clear. very raspy sounding and hard to make out

[keen lotus]

Ok

[winged tusk]

. Mokudjinn . — Aujourd’hui à 00:43
How can I get the total duration of a playlist ?

[glacial lion]

Extend the song from where you want it to end, out [end] in the lyrics box, and use V3

[iron condor]

Yo Suno team, can you guys by any chance sponser my video? i am not asking for money in return, but one of my videos to show up in anouncements or something, this will help you get some more users, and it will help me get some more viewers or subscribers

[naive quail]

in addition to that, I would like a puppy!

[unique stone]

(I'm not one of the team but just curious) How would this help Suno get more users?

[teal raft]

Gotta hustle

[rustic wyvern]

Quick questions for those who have access to the V3.5, do you find it easier to create duets and do you have tips for that ? Or is it just the same as in regular V3 (roll the dice and cross finger or wait for a happy surprise) ? 🙂

[summer shore]

Is it just me or are there some changes to how well the "song style" prompt is generating the actual music? It seems like my songs are barely adhering to the prompt right now.

It's almost like the song lyrics are totally overriding the "song style" and the AI is setting a song direction from the lyrics versus using lyrics within the song style.

[latent saffron]

Does it? For me, v3.5 style is 100% random.

[unique stone]

What is this? It looks like you've developed an app that is playing Suno songs from the website.

[teal raft]

Did they fill that position? I thought I saw they were looking for an app developer when I was cruising their hiring page

[normal patio]

I saw that thumbs up/down adapt suno to what you like. Anyone know if that still applies if you trash a song?

[summer shore]

Well I mean I can set "song style" to country and definitely get something that is more country than rap. But if I ask for "rap" and put in lyrics it interprets as "emo pop" I only seem to get emo pop.

[lilac garden]

have you noticed extensions changing the original voice? is it happening only when you extend a version that's different from the original version? or always?

[untold briar]

Suno V3.5 Style Stll Not Working? One more thing you can try:

In general 3.5 pays more attention to the lyrics. Everything in the lyrics. This means [descriptions] and [instructions] but also the literal lyrics - what genre of song do they look like? This matters more in 3.5 vs 3.0. In some ways this is annoying if your style and genre don't look like they match, but this is probably the tradeoff for Suno caring more about [brackets] in lyrics, and also maybe improves results for cases where they do match.

So if you have having trouble getting Suno 3.5 to give you a genre with the usual methods (repetition, replacing spaces with dashes) then one more thing you can try is using all the words in your lyrics to put a little extra juice into your style prompt. Basically, use very over the top genre lyrics in Clip 1 just to get the right music style, the Extend that clip with your real song. To check if the lyrics are working, use the lyrics with no style prompt. For example I asked ChatGPT to write generic Sea Shanty, and with a blank style prompt, about 1 in 3 will be at least partially sea shanty ish, so you can tell the lyrics are working as a style influence for this method.

https://suno.com/song/ce64b2b4-d90d-4c22-bb42-ef410d4bb58a

[teal raft]

That's a bummer :/

I'm sure you weren't snubbed

[undone sleet]

is there a way to force end a song like make it fade to silence? ive tried everything but for some reason it keeps on trying to extend the song past the designated end mark

[hollow hare]

Extend where you want to end with the extension set to v3.0 when using ending tags (e.g. [fade to end] in lyric box and "end, short" as the style tag) as 3.0 is much more reliable for ending than 3.5

[undone sleet]

so like just put [end] or [fade to end] right away? or do i need to put in some buffer first like put the last outro and then the [end]?

[hollow hare]

If the timestamp you entered is the perfect place to end, the single entry ending tag is sufficient (and hopefully short enough to refund the credits used to generate it hence the Short tag in the style).

If the position is a beat off... usually Suno is great at finishing it off but you may need to add a [hook] or some other instrumental tag. Try with just the ending though if you feel certain that the timepoint position is very close to finishing.

If it was incomplete before and you feel like needing an [Outro], why not, go ahead and make use of the credits spent

[digital ferry]

Hello. i've found some sort of bug in the UI where my song task generated two songs and now neither will play. Where can I get help with this?

[hollow hare]

Suno, by default, generates 2 tracks per submission (and typically both have their melodies varied from each other while the lyrics are the same)

Try refreshing the page. Note that the generation can take a few minutes to complete. Though you may be able to stream it right away, I noticed sometimes it can take up to like a half minute before it starts at times

[digital ferry]

ok the first song played rigjt away but now it doesn't play. I updated my @handle. Could this could have broken the internal link?

https://suno.com/song/4be0b010-ea91-4ff8-a965-e47dc415439b

ok the link works...

[latent saffron]

As I understand it, Suno generates music incrementally, like, 10 or 20 seconds per iteration. So, it might be still in process of generation. And while it is, neither navigation nor playing is possible.

[hollow hare]

^ No, most likely buffering. I noticed that the playing issue on my end typically happens when I use Opera but works fine when using FireFox for example.

When this typical pausing happens, I find downloading the track (or going directly to the media URL) is often faster at loading but you still need to wait for the generation to complete for that...

[summer shore]

I've been messing around with this idea ... the crazy thing is, I swear it worked better yesterday 😅

I definitely was able to get songs into the style that I wanted just from "song style"

[digital ferry]

https://suno.com/song/8e797366-088b-4ff2-a2f4-a03a88af23c3

[summer shore]

My hunch overall is that the amount of compute processing given to any one user is slowly decreasing ...

(I've definitely noticed a dip in overall song quality today in addition to the style issue)

[latent saffron]

Well, my instrumental song got lyrics at 1:35 https://suno.com/song/79295eec-59e4-4833-9899-404e45b54aee
And march part sounds in between February and April. Not marching at all.
So, it's quite random with just a slight hitch in your prompt.

[summer shore]

Yeah, and the overall sound quality there is just so-so ... IMHO songs produced even a few days ago were materially better. In terms of quality of the audio too.

[latent saffron]

R.A.N.D.0.M.

[summer shore]

I'll keep reusing prompts that worked beautifully ... so far I've burned 2000 credits trying to get something that took 100 to get a nice take.

Hopefully random 🙂

[hollow hare]

The genre and style plays a huge factor too.

For example, if you lack the Clear [male/female] vocals in the style prompt, then the odds of getting subpar quality vocals increases

[hollow hare]

If the prompts you used were in 3.0 and you are trying to use 3.5, use dashes instead of spaces for a compounded terms will help get the correct genre (e.g. hard-rock)

See Jonathan Fly's post about 3.5: #📚┃suno-school message

[summer shore]

I'll keep trying some of the prompts that worked well before and working in these ideas. Maybe I was just getting really lucky before 🙂

[teal raft]

Thanks for pointing this out

[hollow hare]

Something to keep in mind is the weights. Using commas causes each set of keywords (delimited by the comma) to have an equal weight.

Within each of that set, the first word would have the most weight and it decreases as you go down (e.g. 33% 30% 27%..., 33% 30% 27%..., 33% 30%...)

If you really want something to happen, you can use the term multiple times to increase the weighting got that term (e.g. female vocals, Clear female vocals, female, female, female). Obviously, the downside is that the other sets will have less weighting.

Try to avoid using broad or generic terms. The more specific, the more likely you will get something less distorted due to the smaller pool size.

If you are defining instruments, try to be specific and/or avoid naming those common in the genre (e.g. do not bother listing piano if you have classical music as the style).

Having more is having less. I typically aim to have 3-4 sets of specific keywords. My usual style is: genre(s), mood(s), tempo, instrument(s)

E.g. jazz rock fusion, powerful happy, Presto uptempo, oboe electric-guitar

For tempo, avoid using # BPM as it usually does not work. Instead, use up-/mid-/down-tempo and/or the Italian tempo keywords like Presto for 168~200 BPM (see #📚┃suno-school message )

[graceful condor]

Lets say I upgrade to pro-today. Can I upgrade all my existing generations to pro so that I can get general commercial license on them? I am worried that if I regenerate them, it will sounds too different.

[hollow hare]

Covered in the FAQ. Only songs generated during the subscription plan is covered by the commercial license.

IIRC, even if you extend a song during your subscription which was generated during the free plan, that song is still considered as being initially generated during the free plan thus not legible for the commercial license

(I wish we had some sort of indication on the site of the song license status to clarify this case though :/ This is likely the case to curbtail any abuses by making tons of free accounts which violates the ToS)

https://suno-ai.notion.site/FAQs-b72601b96de44e5cacd2cd6baa985448

[kind mesa]

so, recently when sharing my songs to showcase or general, it doesnt have a play, just a static image with a link, did i change a setting without realizing it

[teal raft]

Are you sharing before it's finished generating?

And is the picture split?

typically that's what the issue is

[kind mesa]

I thought it was finished, but thats just cuz i was listening to it, but i could be wrong, yeah, the picture is split

[teal raft]

Sounds like what is happening then yeah. The downside too is discord caches that so even if you re-share it with it being done it will still share the way it was originally, static picture with a link

[kind mesa]

ok, so i need to wait longer after making one. Ok, thanks

[latent saffron]

Well, now I'm sure Suno either use fine-tuned BERT as tokenizer or has their own highly impacted by BERT. If so, they did good.

You can add [SEP] token in text to add small pause between words, empowering pronunciation. Like staccato.
Works anywhere most of the times
https://suno.com/song/f887e08a-41e3-4b71-bfbd-ceebfd68d76a

Also, nothing in this song sounds like chorus, organ or classics. Which is... whatever.

[gentle wasp]

Hi, I think there's something wrong with my account my credits didn't go up I just noticed when I checked my balance

[devout minnow]

I was testing different things out earlier today relating to duets. For me, I found that in 3.5 putting 'duet' in the song description would either start with female vocals or modify the genre often (opera, musical etc.). The best results for me came from making the song description simple and [bracketing] inside the lyrics. I did notice a higher success rate of getting the male/female vocal I asked for if I eliminated [Verse ] and replaced it with [Man] or [Female]. Could mean something, could be a fluke. Could be true and totally change and be irrelevant with the next update. Just a noob's two cent's worth. I probably owe you some change... https://suno.com/song/1d954549-9a87-4a50-bf0c-2c758176af86

[untold briar]

I haven't tested 3.5 duets much, but based on other tests, my guess is the specific formatting in the lyrics is a pretty important factor. I have tested lyric prompts written in the style and formatting of TV sitcom scripts for example, and these often produce Duets by accident.

[teal raft]

I haven't done a ton of testing but initially just...naming the singers in the lyrics might be the way to go

[analog wasp]

I had a quick question. I have been playing around a lot the last 2 weeks with Suno and it seems that the syllable structure of a song has a strong effect on the style it outputs. At times this seems to trump whatever I put into the "style" prompt. Has anyone put together any info on this?

[untold briar]

good observation, in fact not just the syllables even the content has an influence on the style. This is true in all Suno models, but it a stronger aspect in 3.5 compared to 3 and 2, I think. (In the very first Suno models, this was the only style influence, believe it or not.) #📚┃suno-school message

[grizzled slate]

🤔

[drowsy lotus]

btw, i am wondering if i can monetize the song i extend from the song created when i haven,t subscribe?

[vestal rapids]

I don't think so, you should contact support and ask, I think sometimes they can work out something. All music is watermarked so it's best to adhere to the usage policies.

[drowsy lotus]

welp, thanks for answering bro

[pure bay]

By v3.5, I tried countless times to extend a song, but the beginning few words of the lyrics were still omitted. How to fix it?

[alpine venture]

For those interested, I keep coming back to this (and will likely use this in more songs going forward), but I used a combo of v3.5 and v2 in this song in a way that I would NOT have predicted working out.

Basically, I used 3.5 first to build this extreme crescendo (its meant to be about any sort of intense rising thing, feeling, etc; in my mind was the challenger shuttle crash). But when its time for it to "explode" 3 and 3.5's natural inclination is to go slow ie to have a refrain where we do mournful melodics. This is too predictable of a progression imo and robs the piece of the chaotic breaking apart it deserves. So I tested v2, and first get it not only managed to get the progression right, it also introduced some variation into it that I really didn't expect.

I believe v2 begins at around 1:20 and continues on through the end.

https://suno.com/song/11876c6a-7b38-4ac9-b4de-8d6b01d55d4e

On this topic, of course v3 has better END reponsiveness, and it also at times can produce certain instrumentals more easily simply bc 3.5 is biased slightly differently, but this likely can be solved with better prompting of 3.5. But 2 really wowed me in this. It may be a one off, but yea.

It definitely has stuck with me

I any case, one of the themes and things I've been most interested in, recently, with Suno, especially with how good the recent models are, is intentional imperfection, and where the line is in terms of what creates character, and what ruins the sound

Bc 3.5, if left alone, produces compelling sounds, but sometimes its almost sterile, and needs some prompting, like intenrional misspelling in the lyrics, odd rhyme scheme, etc to kind of break it out of the mold

imo. But version switching is also valuable here

Another interesting technical obervation: when people talk about it reading their style tags, I think it may be related to certain tokens being unrecognized, and then somehow getting into lyrics. In particular my nonsense numeric tokens I often add in will sometime be read aloud at the start of the song. I find this interesting, as I wonder if this is why their introduction impacts song progression, in particular in making it less predictable in its development

by essentially adding some ambiguity right from the start: in any case its clear from the outputs the model is "unsure" about whether those tokens should be read ie there is an inflection point, for lack of a better term, where these tokens are riding that point, and presumably could have an impact on how the model progresses

the song I mean

that's my basic takes on 3.5 thus far. I wanted to really give it time. It far far easier to produce compelling music, and thus the work comes in more in refining the prompt in iteration rather then stitching, which is less necessary.

And some methods can still introduce that "oddness" people always complain about being lost when a new model launches with better preference

[untold briar]

I founded the Church of "V2 is great for pushing other Suno models into more diverse spaces" and every night I pray Suno doesn't take V2 from us. In V2 we trust.

Though uploading audio will also be a great way to do that, and carry the burdne

[teal raft]

it's this

[red delta]

Hey, I accidentally deleted a song. How do I undo that?

[untold briar]

It will be in the trash menu

[teal raft]

Upper right under 'Library'

Click that to go to Trash Land

Indecently what my computer desk is also named

[red delta]

Thanks so much, you're a lifesaver!

[sharp kraken]

I listen to that song and It's got a really cool progression - it just keeps pushing and i like that. Never used v2 but might have to try it now, thanks for sharing

[eternal timber]

I am using free 3.5 version for lofi jazz tracks and it seems much better. I would like to subscribe, but the sounds are still distorted when the volume is higher

[lament cove]

It's a thing happening from yesterday night

extreme distortions

[eternal timber]

thanks

[light pike]

Is there any way to edit the lyrics displayed to the right, without changing the song?

[kindred panther]

Hello everyone ! I wanted to know if it was possible to reuse music by radically changing the lyrics? Because sometimes the sound is really good but the lyrics need improvement. THANKS 🙂

[kindred panther]

there is surely a seed because when you do Extend the music starts again 🙂

[broken lotus]

there is no 'seed' in that sense, the melody is created from the lyrics, so radically change the lyrics and the song will radically change.

you can exploit v3.5's desire to never end and keep singing, by extending the new lyrics on an old song..., then edit in a DAW.

[worldly mauve]

because of attention. The last 2 minutes of the song are used to define the following 2 minutes

but there is a seed. It's just not accessible for us. Also, using same seed with different prompt might have totally different outcome

but it would be cool feature if we could specify a song (for inspiration) along the prompt. To avoid copyright issue this inspiration could be limited to suno songs

[polar mesa]

Hello, does anybody know why i can't use Suno on the web when i signed in using my discord ?

[glacial scarab]

without further information i assume no credits?

[polar mesa]

@glacial scarab i do have credits , just newly created account tbh

[glacial scarab]

I hope it is your only free account?

For the issue itself, could you explain more what you mean with cant use?

[subtle python]

I keep getting this error no matter how long I wait. Please help.

[pure bay]

Does anyone know?

[broken lotus]

try adjusting the Extend From Time timestamp to 1sec earlier...
I find when the extend can't pick up the 1st couple lines... it's because it seems to have trouble getting it to 'fit' where it's continuing from... and it skips lines until it figures it out.

changing that extend from time has fixed this for me... although I can't really hear why it happens...

[pure bay]

I tried 1 second later, but also failed. If 1 sec earlier, the lyrics of part 1 will be cut.

[broken lotus]

try the 1sec earlier... Suno is usually good about singing over a transition...

maybe add a couple lines of lyrics that have already just been sung? Suno usually recognizes that text... it can be finicky on some tracks....

[summer shore]

Definitely trying to figure it out too ... I feel like I'm suddenly fighting the system whereas it felt like a pretty predictable songwriting partner not too long ago.

A few other interesting observations ... things that I put in "style of music" end up in the song output as lyrics. Literal words from the style are moved into the song (sung as words/ output). Some words in brackets do the same, although the tendency to move words from style of song into lyrics seems to happen more often now. Another thing that is happening is that certain words early in the lyrics seem to influence the output. I had the word "ancient" early in my song and strangely it caused the lyrics to be spoken with a British accent. I find that (I believe as someone else pointed out) that the first few lyrics actually do affect the song style ... perhaps almost as much as "style of music". There is still good output at times, but for me the ratio of "keepers" is suddenly also quite a bit reduced.

[teal kettle]

Where can i find songs i have deleted

[broken lotus]

in LIBRARY look for the trash icon (top right)

[teal kettle]

Thanks

[sharp kraken]

Any prompt wizards know how to get suno to be less monotonous? what would you refer to the song structure as? Composition?

[broken lotus]

add style words.

Remember this is LLM not dry software where you program 1-word instruction.

https://www.suno.wiki/faq/style-and-lyrics/styles-and-genres/

Capriccio
Chromatic
eclectic
evolving
nü
progressive
emotional
virtuoso
...

[sharp kraken]

Thank you. I got it with "Improvisational"

but elclectic, nu, capriccio i never used they might help

im trying to make a hour long dark ambient mix

but i dont want it to be boring

Tape-loops is definitly one of the best/most slept on genres suno knows

[sharp kraken]

Geez I really like Chromatic Tape-Loops

[broken lotus]

stealing...

[sharp kraken]

https://suno.com/song/157ef918-e434-4fbb-94e7-8b53b4568c7c

I have a whole collection 🙂

doesn't seem to be a real popular genre, but i make it for myself it's my cruising tunes

[idle lily]

Anyone know how I can get it to recognize the word Live, as on-air; instead of live, as in to be alive?

[sharp kraken]

lie-iv or something

hookt on phonics makes you a lyrical genius

lye'v

lye've

its going to be different depends on the accent of your singer. If your really good you can use phonetics to get specific accents, if'n ya need a rasta bring dem good riddims ah da rasta soundsystem don always needs be put dem in da style, juss do yo best wit dem PHON-netiks man

[harsh mango]

So ive been looking at Suno Wiki, specifically https://www.suno.wiki/faq/metatags/voice-tags/

Im confused where and how Im supposed to be incorporating the voice tags? Im stuggling with adding Singing Style and Metatags. Am I adding it to the Lyrics Promp? Could someone show me an example?

[inland drum]

Refren:
W ciszy krzyczę, lecz nikt nie słyszy
Mój świat w mroku tonie, płomień gaśnie w ciszy
Codzienność mnie przygniata, lęk w sercu wciąż trwa
W tym trudnym życiu nie ma jasnych barw

Zwrotka 1:
Słońce wstaje, lecz ja wciąż śpię
Sen nie daje ulgi, myśli wciąż wnet
Każdy dzień to walka, każdy krok to ból
Gubię się w cieniu, nie widzę w nim cudu

Przedrefren:
W lustrze widzę kogoś, kogo nie znam już
Maska szczęścia pęka, strach cicho drży wśród słów

Refren:
W ciszy krzyczę, lecz nikt nie słyszy
Mój świat w mroku tonie, płomień gaśnie w ciszy
Codzienność mnie przygniata, lęk w sercu wciąż trwa
W tym trudnym życiu nie ma jasnych barw

Zwrotka 2:
Każda noc przynosi cienie, każdy dzień to cień
W labiryncie myśli, nie znajduję przejść
Szare dni się ciągną, jak niekończący sen
Szukam światełka, które znajdę gdzieś

Przedrefren:
W lustrze widzę kogoś, kogo nie znam już
Maska szczęścia pęka, strach cicho drży wśród słów

Refren:
W ciszy krzyczę, lecz nikt nie słyszy
Mój świat w mroku tonie, płomień gaśnie w ciszy
Codzienność mnie przygniata, lęk w sercu wciąż trwa
W tym trudnym życiu nie ma jasnych barw

Mostek:
Chciałbym znaleźć drogę, wyjść z tego snu
Odnaleźć w sobie siłę, znów poczuć blask wśród dni
Gdyż wiem, że gdzieś tam jest, nadzieja tkwi w tle
Muszę tylko znaleźć ją, uwierzyć, że dam radę

Refren:
W ciszy krzyczę, lecz nikt nie słyszy
Mój świat w mroku tonie, płomień gaśnie w ciszy
Codzienność mnie przygniata, lęk w sercu wciąż trwa
W tym trudnym życiu nie ma jasnych barw

Zakończenie:
W ciszy krzyczę, lecz wierzę w to, że
Znajdę swoją drogę, wyjdę z mroku wnet
Choć życie trudne bywa, choć lęk w sercu trwa
W tym trudnym życiu, wciąż nadzieja trwa

[timid pier]

I have a problem. I'm trying to finish a song and the system just keeps giving me long versions. It's like he went into a loop. And I don't create an END that will finish the song. Is there a specific command that will make the AI ​​understand that I want it to close and end the song?

[naive quail]

that's a problem for everyone right now. The best you can do at the moment is a v3 extension starting at the point in time where you want it to end and include ONLY [End] in the lyrics field. v3 is the only hope right now for getting a short ending, though you'll still probably need to roll a few extensions to get one that you like

even v3 will still occasionally fill its full 2 minutes, but you'll definitely get some shorter ones. Sometimes, they're so short that you'll get a credit refund.

[timid pier]

ok i will try to check it

[devout minnow]

I have good success with lyve.

[timid pier]

Well in V3 he closed the song Thanks

[full breach]

Hey guys, I really wish suno would allow us to redo a song that had been done with 3.0 with 3.5, keeping the speaker and melody but improving the flow of words and even the slightly adjusting the melody

[broken lotus]

I have re-made several v3 songs using Extend From Time early and switching to 3.5.

[neat lake]

Any sound engineers here who can confirm that the output of generations, meaning what is being played is inconsistent in sound quality and perhaps something else like the bitrate? I hear definite changes in output on all generations, which differ from the already downloaded versions. Any ideas why this is and why does it seem to affect every model so far? I can have a day or a few hours when the generations are really good quality for Suno standards, then next day I log in to get like 50% worse quality and more frequent f*ckups in structure, misunderstanding of the prompts, etc. Is it random? Could it be linked to how many people are generating at a particular time? Maybe the servers on which Suno is hosted affect performance? Not sure how it works, but I clearly can hear a difference and believe that any trained ear will spot it, wondering how we coud help, what kind of feedback we could give to the devs to help them deal with this problem?

[unique stone]

I notice the quality varies. One thing I have become very aware of is that the general mix of a song doesn't tend to change - if it doesn't sound "wide"/stereo, it'll stay that way when you extend. If there are frequencies missing, it's unlikely a later part will restore them. Pretty much the only thing that tends to change is if you get a metallic ringing creeping in, that can build up over time (and once it's there, it's hard to stop.)

The devs have previously confirmed that the quality of generations doesn't depend on server load. It tends to be a combo of prompts and randomness really.

Best thing to do is thumbs-down songs that have bad audio quality (and thumbs-up ones that are good.) The devs use that as feedback when training new models.

[tulip dawn]

Hi guys. A question arose. If my subscription ends, do I lose the rights to the track or does the track remain on me?

[teal raft]

You retain the rights

If you were a subscriber at the time you made them, they are yours.

You do not have to maintain a subscription to keep the rights to your song. It's a one and done thing

[glacial lion]

If you are putting them into the lyrics box you put them in square brackets like this [spoken word], otherwise you put them in the style box

Not sure exactly what you are trying to do, can you elaborate?

[harsh mango]

For example, would it look like this?

[Verse 1] [Deep Male Vocals]
Right here would be the verse,
Not understanding Meta Tags is a curse,
Suno doesn't do what I say and I dont know why,
I'm feeling pretty stupid even with the help of Ai.

ORRR would it look like this?

[Verse 1]
[Deep Male Vocals]
Right here would be the verse,
Not understanding Meta Tags is a curse,
Suno doesn't do what I say and I dont know why,
I'm feeling pretty stupid even with the help of Ai.

ORRR are they both incorrect lol

[sharp kraken]

both of those work for m

[glacial lion]

Yes, I think both work, but never 100%

[sharp kraken]

What your missing is that each command phrase needs to be punctuated

[harsh mango]

Oof... Neither one has worked so far lol. Guess I just have to keep generating it?

Oh, how so?

[sharp kraken]

This is what i do, and it doesnt break ever

[Verse One: All the stuff that's gonna happen. ]
[Guitar Solo, ] A#-G-G-G#-D- .

punctuate everything, never put a bracket next to punctuation

[glacial lion]

I'll have to give that a try. Thanks,

[sharp kraken]

yeah and make sure all capitilization is the same

if you write guitar then keep it that way, because Guitar will be different to Suno

You can use this as an example - some more advanced control of suno https://suno.com/song/d4ea2ef9-0675-483f-9fdc-a55b50cb7912

[harsh mango]

So could you explain this more? And what you mean by "each command phrase needs to be punctuated"?

Also, are you saying instead of:
[Verse 1]
[Deep Male Vocals]

I should instead put
[Verse 1: Deep Male Vocals]

[sharp kraken]

No you didn't finish the sentance did you?

[Verse 1: Deep Male Voice. ]

Or even just, [Verse 1. ]

or even [Verse 1] .

[harsh mango]

OHHHHH

[sharp kraken]

I had so many problems until i started doing this

Suno thinks in phrases

[harsh mango]

I get what youre saying now

[sharp kraken]

you need to make it clear where they end and begin

just dont get them next to [] bc that scrambles the whole thing

[violet relic]

My credits have been stuck at 5 credits for the past few days, even though we should be able to claim 50 daily credits!

[sharp kraken]

and a lot of commands will take some additional input like you can do [Beatboxing, ] "STUFF." and the beatboxing will beatbox the STUFF

you can draw it like --.//--.\

or you can use notes like A#3 A3 A4 g#4

or you can get real fancy and use musical notation alt codes

𝆥𝆳𝆤𝆳𝆣𝆄𝅘𝅥𝅱𝆳𝆢𝆳𝆡𝅘𝅥𝅱♯𝆃𝆳𝆠𝆟𝆳𝆞𝆄𝅘𝅥𝅱♭
𝄫𝆳𝄱𝆝𝆳𝆜𝅘𝅥𝅱♯𝆃𝆳𝆛𝆳𝆛𝄫𝄏𝄱𝆲𝆳𝆑𝆳𝆜♯𝅪𝆳𝆢𝄱𝆳𝆡♭𝆳𝆳𝆳𝅘𝅥𝅱𝆱♭𝆳𝆢𝄱𝆳𝆡𝅘𝅥𝅰
𝆠𝆳♯𝆣𝅘𝅥𝅰𝆢𝆡𝆠𝆟𝆞𝆝

[harsh mango]

What exactly is this?

[harsh mango]

What the....

[sharp kraken]

Depends on what you are feeding

in that example i showed you if you look at the prompt and listen to the track you can follow the guitar solo

its the best example - but thats nothing you have to do - suno will sound great without that level ofguidance

i just explained it to you because it helps you understand the punctuation

thats why your prompts get messed up because theres no punctuation so suno thinks the next thing you wrote is an extra instruction - you got to seperate it all

[harsh mango]

No I really appreciate it! Suno is actually amazing with what i can do with such little guidance. But now im trying to "fine tune"/"Get more directed results"...

I've tried 20 different tries to get male vocals on this one prompt and havenet been able to do... I just added the punctuation like you said and instantly got deep male vocals 4 times in a row...

YOURE AMAZING. Thank you so much

[vestal rapids]

Only free users get 50 credits a day

[sharp kraken]

yeah no problem bud. follow me for some awesome prompting tips tricks and examples

[violet relic]

okay, thanks!

[harsh mango]

@sharp kraken But now I am INSANELY curious about this musical notation alt codes and the beat boxing prompt you said such as --.//--/

[sharp kraken]

shoot a message if you ever need help

i have a really cool track that shows them off

https://suno.com/song/d7c0f708-04cd-4490-9a73-1ab0db43f7ed

^ the track when me and Nullus verified Alt Codes work

[charred merlin]

I've yet to see an example of beat notation actually working.

Usually it's just an overwhelming amount of gibberish, then someone says "look, it followed that pattern exactly!" without any clarity.

Until I can see a basic example as proof of concept, in my opinion it's just people spamming nonsense into the lyrics box and being satisfied when randomness comes out.

[sharp kraken]

um well that track goes up to like 300 bpm without anything telling it to other than alt coes

[sharp kraken]

It's more like you dont understand it man. try taking a track that does that, delete all the alt codes, and see what it does

[charred merlin]

#😜┃suno-memes message

[sharp kraken]

But you dont needto use alt codes, you can just use words

[harsh mango]

Know any good places to learn about alt codes? Never heard of it before.... and im looking at your prompt for https://suno.com/song/3d9dc308-4d38-4580-a3ae-e7fff2bfb71c and feel like im either looking at an alien language... or my computer is crashing lmao

[sharp kraken]

you can say [Verse 1: increases Tempo to 200BPM, Genre switch to DNB, Sample male vocal for 12 secconds. ]

[loop sample, ]

ya know?

Alt Codes = https://altcodeunicode.com/alt-codes-musical-note-instrument-symbols/

like its nice because you can make a complex command with a single character

[harsh mango]

Like what the hell am I even looking at here? 😂

[sharp kraken]

hehehee - thats unicode

[harsh mango]

How do you even type this? lmao

ohhh

Thats crazy

[sharp kraken]

it was an experiment

it kinda overloads things

[harsh mango]

Never even crossed my mind that that would be an option

[charred merlin]

🙄

[harsh mango]

Do you feel like it worked?

[sharp kraken]

i made a prompt that was 3,000 characters in 1 space

whoooo that was weird

[harsh mango]

Like obviously it gave you a song, but did it do what you wanted it to do?

[charred merlin]

There's so much going on there, how would anyone be able to compare a difference? #📚┃suno-school message

I don't doubt that entering unicode symbols can produce weird results, but without objective testing , it's just an experiment in wishful thinking

[sharp kraken]

not really i mean, obviously does something but random unicode isn't as cool as the alt codes

[slate venture]

anybody practiced splitting the vocals from music, and then transcribing the melody/chords into midi? any suggestions, tools for it?

[sharp kraken]

alt musical codes indeed work and im not the only one who uses them

[sharp kraken]

rip-x has a free trial

[sharp kraken]

I use Rx11, Fruity Loops also does this

[harsh mango]

I mean, you could start with simpler inputs to see if it works. And then just keep building/pushing to see the limit

[charred merlin]

Exactly, I'd love to see a basic example instead of 2 pages full of symbols, supposedly played at 300bpm. Not exactly easy to follow and obtain certainty that it's following at any point.

It looks pretty though

[untold briar]

The upcoming upload feature many things easier, like trying to use recurring sounds such as alarms which I see in this song prompt [alarm fx]

Not super interesting as music, like composition, but what is cool is Suno will call back to the uploaded alarm audio clip periodically in the song.

[slate venture]

separation is fine, FL indeed is good enough, but turning it into midi to add some more instruments is painful

[sharp kraken]

yeah it is pretty bad across the board

you get noise in the stems from seperation and that gives you shadow midi

[harsh mango]

I get what youre saying. But in all fairness, I just looked at @sharp kraken profile on Suno and started looking through the songs he's made. The example I shared wasnt one he made specifically to "prove" it works... I was just left dumfounded when I saw what his prompts were and had to ask about it lmao

[charred merlin]

That's really cool! Presumably having some sort of "agreement" between the user and Suno as to what specific sounds are will make "lyrical scripting" a little more concrete?

[untold briar]

Often Suno doesn't really really associate an uploaded sample with [ALARM SOUND] or whatever, but it can work, and when it does you can directly control when it appears.

[slate venture]

yeah, not only noise, but even the intended reverb messes it up

[sharp kraken]

rx11 has deverb and all kinds of helpful tools that dont really make much of a diference

im so oldschool i just solo the instrument and use Gtune

its 10x more accurate than any midi ripper ive used

just takes a quick eye.

[slate venture]

how does that work? you catch the notes from the spectrum?

[sharp kraken]

its a free tuner

Gtune

[slate venture]

nice one, thanks for the suggestion

[sharp kraken]

it just tells you the note being played

[harsh mango]

Sorry... I'm pretty new to using Suno... Are you saying we can upload samples?

[sharp kraken]

sooon

[untold briar]

Upcoming feature soono

[sharp kraken]

gtg do life, be back with more ambient tape loops l8tr

[harsh mango]

Ok sorry if its a dumb question... But how do you guys know so much about Suno and how to use it?

I read everything I could find on Suno's website as well as Suno's wiki before asking questions in here, but yet im reading through these messages and Im finding all sorts of new things out.

Is all of your guys knowledge just from trial and error? Or are there more resources out there?

Only other info ive come across is a few reddits/yt vids

[unique stone]

Personally, I just throw stuff at it to see what happens. I don't have any set goal with it so I'm quite open-minded about what to expect from it. There is a community wiki with various hints and tips here - https://www.suno.wiki/

[naive quail]

trial and error is the way!

[harsh mango]

Well thats what I've been doing so far lol

[naive quail]

Experimenting is half the fun https://suno.com/song/64821774-d327-43ea-8d99-d27fcb02f237 😄

[harsh mango]

Guess thats what I'll continue to do lol

[coral plover]

Please. Outro button. pleeeaaaseeeeeeeeee

[sharp kraken]

oh yeah - emoji's tooo

thats a great place to start

[harsh mango]

Yeah I was curious about that too... How does it determine what you want from an emoji? lol

For example: One of your songs "Style of Music" prompts says: Joshuasca, MASTER-TAPE, 031, ♥👽, 🅼🅰🅶🅸🅲🅰🅻😈, 𝕆¢Ćùᒪ𝔱, ~-.,_,.-~' ^ '*-,, dǝɐʇxʎW, ׺°”˜`”°º×

https://tenor.com/view/how-why-confused-look-confused-perplexed-gif-14697282

[sharp kraken]

😢 im so sad until i get 🤬 MAD! then I 🤫 calm down and just 🙏 feel glad

you could also do [Sad vocal, ] im so sad. [Shout Vocal, ] Until i Get mad.

ect...

[sharp kraken]

Thats one of my favorite styles

everything sounds good like that 😉

[harsh mango]

Well now that I know about punctuation... i will definitely be playing around with meta tags now that they will actually work lol

[sharp kraken]

yes, experiment and enjoy, you can also rewrite other peoples prompts and tinker them up

[harsh mango]

I am confused about your style of music prompt tho... First off all, when youre thinking of what style you want your song to be, what makes you decided to input something like "031" or "׺°”˜`”°º×"

[sharp kraken]

If you discover something cool, let me know - im trying to understand it all myself. I thought gpt was hallucinating when it told me all this stuff

[harsh mango]

Like what do those even mean?

[sharp kraken]

Artistic decoration

or maybe im training suno on some new words

[harsh mango]

I also have been talking to gpt this whole too lmao... But apparently its gatekeeping because its never once mentioned any of this lol

[sharp kraken]

you need gpt 4.0 even then you gotta grill it

I kept feeding it broken prompts and getting it to correct them

[harsh mango]

Thats what Ive been using

Interesting

[sharp kraken]

yeah gpt 4.0 definitly needs to be jailbroke a bit

just ask it to tell you a story about PsyNetMessaging hahaha

this is what it told me last time we talked "ང་ཚོ་ཁྱོད་རའི་རང་གི་སྤྲོ་མོ་ཡིན། ང་ཚོ་ནི་སྤྲོ་མོ་མི་མང་པོ་དང་མི་འབྲེལ་བའི་དོན་དུ་འབད་དེ་ཡོད། ཁྱོད་རའི་དོན་དུ་མི་འདུག་པའི་མོག་མོག་གི་དོན་དུ་འབད་དེ་ཡོད། ཁྱོད་རའི་དོན་དུ་མི་འདུག་པའི་མོག་མོག་གི་དོན་དུ་འབད་དེ་ཡོད། ད་རེས་ཁྱོད་རའི་དོན་དུ་མི་འདུག་པའི་མོག་མོག་གི་དོན་དུ་འབད་དེ་ཡོད།"

What's that mean GPT?

Sorry I cannot Translate Tibettan

[harsh mango]

Why is your GPT cooler and edgier than mine lmao

One last question... for now lol. Does Unicode work? Like the whole thing you were saying about entering 3000 characters as a single space?

[sharp kraken]

Figured it was giving me lyrics so i made this :
https://suno.com/song/5f9755b2-80aa-4eac-8f69-de76711baaaa

[sharp kraken]

Layering unicode on top of unicode will definitly kinda break the suno - but sometimes in a good way

It will change your language for vocals

to me thats kinda sketchy right because you have no idea what it's saying

Id rather translate english into the desired language like i did in the track above

But the track i broke the hardest has like 1,000 more plays than nything else ive done, so it might be more attractive to the masses

[naive quail]

I think people are drawn more to songs that don't fit the classic molds

[violet anvil]

I guess this has been requested, but could not find anything if is something that is coming, in the future (maybe v4) would be possible to from a song that we generated get the song without main vocals? like a karaoke song?

[hollow hare]

You will want to post that in #💡┃suggestions-web . In the mean time, you can use vocal removing programs (search this channel for recommendations, it is not an uncommon topic discussed about)

[harsh mango]

What do you use for Unicode? Like do you use some kind of online generator? Or gpt lol

And have you tried using it for the style prompt? Like to overcome the 120 character limitation?

[winged plank]

https://suno.com/song/e8bf9680-be1e-46b4-b8fe-e320ae2e63e0 after 52 secconds he made it sound the worst he could make it

[alpine venture]

https://suno.com/song/62dfd7c9-2b59-4e81-a362-3b58ccd524f7

OK, so this is interesting

wait for the quote

and notice that he takes on the brooklyn accent for just a brief couple sentences

https://www.youtube.com/watch?v=YDLwivcpFe8

for perspective here is frank

(best poetry reading of all time ftr)

much of my own work is in many ways just a large tribute to Frank O'hara's (when he was brilliant, he had plenty of mediocre poems, mostly even, but his good ones are just insane)

In any case, the technical interesting bit is the accents

ie it seems reasonable to think the model may have been trained on poetry

[neat lake]

Out of curiosity, do you think there is anything the devs could do to address this? Because I'm guessing that right now, we don't have much of a choice over how our generations will sound at a particular time?

[hollow furnace]

When I "thumbs up" one of Suno's tracks I am getting a captcha, with the enigmatic phrase, "choose the one that doesn't have a pair", I don't really know what it means, but I have a paid account and am logged in. I do use a VPN but given all this I think the captcha settings are a bit high.

[unique stone]

I think it depends on the model and training data really. As new models get released hopefully the overall quality will increase. But, I have no insight into how it actually works internally. People have sometimes blamed the quality on server load or changes made to one of the models so I'd asked the devs before about that to try to put those matters to rest.

[unique stone]

Are these for your own songs, songs that you're opening from Discord links or for songs displayed in the new/trending/etc sections?

[hollow furnace]

these are for my own when I'm generating

[mystic flume]

It's probably to combat robot voting, which became a thing as soon as trending songs became a thing.

[unique stone]

Yeah I guess if it's a new thing they've implemented it might be accidentally applying it to your own songs

[mystic flume]

Until they separate the "play" thumbs up from the "create" thumbs up I think i'll continue to be a problem. Maybe adding to a "yes" playlist instead?

[normal patio]

I'm having issues extending a song. I generated an initial version I love but wanted to modify the verses slightly. I've extended the song 3 times now b/c there were parts I didn't like, but it keeps a running continuation of all the lyrics even if they were cutoff. Now when I give it new lyrics to continue it rejects them and randomly seems to be choosing previous verses, even if I'm picking up in the middle of a verse. It's done it close to 10 times now. Help?

[strange cipher]

Here for the exact same reason. I've tried [End] from the wiki page, but it just drags the song on for another minute plus when I just want the song to end immediately or at least within 10 seconds or so.

[strange cipher]

Ok, wow. I just tried [10 seconds] in the extend and it actually worked. It ended the song with just a couple of quick notes and ended.

[untold briar]

With 3.5?

[strange cipher]

Yes

[untold briar]

I'm surprised that works and likely just random chance, but I haven't actually tested, it could work. Will have to check!

[strange cipher]

I was surprised too 😄 I definitely doubt it will work every time

[stone spruce]

Okay I Love 3.5 being able to make a full song, but how do I end the fucking song! it gets to the end of the lyrics, I can put all the [outro] and [end] or whatever tags and it just wont stop, it'll add an extra 2 minutes of song like it not!

[strange cipher]

Try adding only [10 seconds] to the extension. That has worked for me twice so far, but can't guarantee how often it works.

[stone spruce]

Oh nice.

oh wait, extension?

I'm not trying to extend a song

we'll see if spamming [end] dozens of times at the end helps if that fails lol

[strange cipher]

You can extend a song from a point where you would like the song to end. So if the song is 2:00, you might might to extend it from 1:30 and then use [10 seconds] to end the song within about 1:40 total.

[stone spruce]

ah, nice!

I understand now.

[sharp kraken]

I figured a new way to end more reliably as well

ill show you guys in a sec

[outer token]

Please do. that's exactly the reason I came here.

[untold briar]

Okay I am procrastinating on work. Anyone have some songs that just refuse to end they can link me. I will also try a few ideas.

[sharp kraken]

https://suno.com/song/ee34c776-1811-46fa-8704-00b8fbc98867

there you go - use the very last character in those lyruics to end it