📚┃suno-school

[sharp kraken]

this is the 1st verse on a 3.5

[strange cipher]

𝇟 <- this character?

How on Earth did you discover that? Is that documented somewhere?

[sharp kraken]

NO

I am just very resourceful and i guess those music lessons when i was a kid paid off

actually now it is documented in this track for all yall - pass it around

its pretty cool to summon instruments for 1 character as well

and dont forget I am writting the rhythm by hand here

thats pretty fun too

[strange cipher]

Thanks 🙂 Can you get it working consistently?

[untold briar]

Well, that's a pretty Jonathan-styled lyrics prompt. Unfortunately it can be hard to generalize from results in prompts like that . "Your songs may vary." 😄

[sharp kraken]

it ended almost all

here

look

[outer token]

Tried it as an extension, and no it did not for me

[untold briar]

There's some different things people mean by endings too. Some people want an ending of the song musically, but it might be 30 or 45s. Other people are like, the song already ended, just STOP NOW.

[outer token]

I'll try it next time I do a full track

[sharp kraken]

https://suno.com/song/53dc711c-c61a-43c6-9549-e3ff0df2e28e

https://suno.com/song/039c5c3b-26e1-4d18-b502-b9259b684605

https://suno.com/song/ff93a391-645c-4594-a466-d0a775dc9e5a

works TOOO good

most of them ended the momment they got there, which sounds weird - so make sure its at the very end

WHEN DO I GET TO WORK FOR SUNO? Your boy needs a JOB over heree

HAHAHA

[strange cipher]

I think the 𝇟 works well for full songs, but I couldn't get it working for extensions of songs. For that, I use [10 seconds] which works fairly consistently for me. I usually get a token refund since the output is so shor too.

[sharp kraken]

well g'night guys, I hope this is useful info - like i spend a lot of my time doing this and I only do it to understand for myself and its something I just freely share even tho I through alot of credits at it. I am always amazed at how deep suno can be so thanks to everyone who made this beautiful beast.

not bad for a homeless veteran with brain damage tho.

[strange cipher]

Thank you!

[stoic stag]

I got a song at over 6 minutes and I've spent over 2k credits trying to get it to work, but at this point, it's just constantly ignoring any and every prompt. It's like it's completely ignoring everything I try to tell it to do and even the lyrics half the time are wrong. It's like I broke it... Any suggestions or is this song kinda screwed. I've been working on this single song for like 2 months now and put so much time and effort into it lels... I got the last verse and outro left... I'm so close...

[untold briar]

Usually best option is to keep changing the Extend From times, slowly backing up further and further. Try and find an earlier spot where you don't have the problem. If you have never changed the time from the end of the clip you may only need to backup 5 or 10 seconds actually.

[stoic stag]

ye i've tried that. i've even tried to extend with a new instrumental only to extend it to try and force it to change, but it didn't rly work all to well. I've tried multiple seconds in like a 10 second time frame, but mostly 30, 31 and 35 second mark in the extensions for both the original and the instrumental extension I added onto it. It just keeps wanting to make it intense and bassy. I think I broke it by forcing it to be intense and bassy in the earlier parts which was completely different from the earlier parts of the song.

[untold briar]

You can Extend from the last clip, or you can "Get Whole Song" on the last clip (or second to last) and Extend from Full Song up to that point. If you have only tried one of those options, could try the other one instead. Other than that, you might be very unlucky and just have to backup even further. It can happen, usually not too often though.

[stoic stag]

ive been extending from every "get whole song". i got... a lot of extensions...

[untold briar]

How long are the clips at the end of the song. The last one, and the clip before the last one. If you can remember or have the links open.

[untold briar]

It just keeps wanting to make it intense and bassy. I think I broke it by forcing it to be intense and bassy in the earlier parts which was completely different from the earlier parts of the song.

I can't know without seeing the song, but if this is the problem, then you may be able to avoid it by Extending from the last clip (not the Get Whole Song "Full Song") instead. If the last clip does not have the intense music in it, then the AI will not know about the intense music.

[stoic stag]

this is the last extension. https://suno.com/song/b2fdb43a-dffa-4119-8cc4-87992e03f5d0
this is the full song i completed with that extension https://suno.com/song/3b954159-c459-414d-8ca3-040efe279b83
then this is the extension with the instrumental to try and fix it, which took me soooo many credits cause even the instrumentals wanted to be bassy and this was the best option out of all of them...https://suno.com/song/8ce72531-702e-4ab9-89da-6b18d2684cda
then this is the full song with the instrumental extension that I'm trying to extend from right nowhttps://suno.com/song/c4466a99-62e3-4614-b8fd-cd8d1e3ca815
I've tried 6:30 and 6:35 mostly.

I've tried so many weird or complicated prompts, but i've always found best success with just simple ones.

[untold briar]

From the clip you called "This is the last extension" - what is the music you want to generate after that? You said the instrumental is to bassy, do you want a less bassy instrumental, if you could pick?

[stoic stag]

basically i tried to get it to be more like the earlier parts of the song, but any prompt I used to try and produce that, was completely ignored so i just assumed suno was incapable of doing that.

[untold briar]

So on this song https://suno.com/song/b2fdb43a-dffa-4119-8cc4-87992e03f5d0 after the lyrics there is a bass heavy instrumental. You DO want to keep the bass heavy instrumental. But then you want to go back to the earlier style of the song after. So you extended that clip from about 30s after the instrumental, but the heavy bass keeps coming back. Is that all correct?

[candid thunder]

Is it available to download separate channels for each song? example (bass, guitar, vocals) in high quality?

[stoic stag]

i didnt extend from that shorter clip, but from the full song version of it, but on that shorter clip, i did it at the 20 and 25 second mark i believe.

i cant really go sooner than that, which is why i tried to use an extended instrumental to force it to change into a non bassy continuation.

i mean the 15 and 19/20 second mark, my mistake sorry.

[clear raven]

problem feedback: I use custom mode and v3.5 to generate a jazz song, but the song doesn't sound jazz, mostly pop

[untold briar]

Are there going to be lyrics after that shorter clip I linked? If there was no bassy problem, what would the prompt be in the next clip?

[stoic stag]

this is like the best version out of a lot that, but the vocals are all robotic and stuff soo... https://suno.com/song/68295522-b93a-4e6b-9676-f01c0209b749

If the vocals werent so messed up/robotic/sped up, i woulda just bit the bullet and went with this clip.

[untold briar]

So if everything worked, you would want to prompt that clip in message above, right after the lyrics end in the short slip, around 15 second, after the words "I'll show that our blazing hearts are IN-DOM-IT-ABLE!!!!!!!!!!"

[stoic stag]

so just delete any prompt to try and manipulate it, even [verse] since its a new verse and just continue it with just the lyrics?

i rly feel like i broke it hard with the 2 previous [chorus] prompts which was just 1 long chorus, but i kept having to force it to try and stay intense through it, which didn't work the entire time, but eh...

[untold briar]

I think there's two problems. Vocal getting worse with a long song, a general problem, and too heavy bass. It's hard to fix the vocals, without changing the vocals.

At least without getting lucky.

[stoic stag]

wow this sounds different with everything before doing what you said. so this is DEFINITELY progress. thx for ur time, i rly, rly appreciate it.

[untold briar]

What was the specific change, using the shorter clip?

[stoic stag]

yeah extending from the shorter clip. never actually done that before.

removing all the prompts and just continuign with the new lyrics as well

[untold briar]

Great. Yeah it took me so many messages to just understand the goal, confusing with all the songs not showing which clip they are for example.

[stoic stag]

and yeah, i think im just gonna have to deal with how much the voice has changed for this last verse and outro.

[untold briar]

You could even try to reset the voice to be clear, but it does probably burn a lot of credits.

[stoic stag]

ah yeah, sorry for linking things so confusingly.... largely my fault there...

i got 3205 credits left lels...

[untold briar]

If you don't mind the voice sounding more like a different singer though, it is easier

[stoic stag]

wasted 2k on trying to fix this so far, well, over 2k but eh...

[untold briar]

I was going to try and fix it for you, but it took me so long just to understand what you wanted, lol.

[stoic stag]

yeah, i rly like the way the earlier voice ended the words and the way it just had its own tone and style.

[untold briar]

Also, oh, if you used my clip, it wouldn't qualify for the contest, if that matters

[stoic stag]

if only i wasnt so terrible at explaining things...

contest? O.o

[untold briar]

The Suno Summer contest. The song has to be all one user, so you could not use a clip in the song if I created it.

[stoic stag]

ive been so engrossed over this one song, i've kinda ignored everything else suno has been doing for da past 2 months aside from switching to v3.5 halfway thro which i think also screwed the voice over.

Hmmmm... i wrote these lyrics years and years ago so i wrote them all from scratch not using AI, but hmmm... maybe i should try to enter once i finish it. never thought I'd be able to actually create my songs i wrote all those years ago since vocalists arent exactly cheap.

[untold briar]

In this specific case, there could be an easy way to clear the voice. If that style prompt, in a new song, can generate a very similar voice in a fresh Clip 1 of a new song. That means you could Extend a clip in your song that does not have any voice, from a very short time early in the clip, and sort of reset the voice back to the original one.

[stoic stag]

so basically, extend from the clip using just instrumentals? every instrumental extension i tried to use would put in vocals like 5, 10 or 15 seconds into it and it was for every generation. hmmmmmm

[untold briar]

Then when you have a small audio clip of a new (clear sounding) voice. Get Whole Song on that clip. This puts the music back in the memory. Then Extend From a time with a small sample of the clear sounding voice, and with luck, it can resume the original song, but with a clear voice. This is like, super annoying and fiddly though. And I think won't be needed with later features.

You do want to prompt lyrics, usually just something generic that isn't a verse, like Ahhh Ohhhh

kind of a mess to explain in text while I'm this sleepy, maybe save for another day, lol

If you check Discord for how people change the music genre of a song, it is similar. By using a very short clip of a few seconds only, you can do a musical reset. This can be a genre but also a voice.

[stoic stag]

isnt that what i did with the i think it was 5 second instrumental extension? O.o

[untold briar]

It doesn't have to be instrumental, and 5 seconds is actually too long sometimes. But the important part is that you extend from a 5 second clip by itself, not a full song. And in THAT next clip (after the 5 second one) you can prompt for a nice clear voice, or a style change. Because there is no voice or music style in that tiny clip, or there is only a little bit of one. Then Get Whole Song after that.

[stoic stag]

OH, i get it now.

is there any specific prompt for a clear voice or just [clear voice] ?

[untold briar]

Just try using the prompt that gave you the voice originally. Actually if the first clip is 3.0, this may not work as well.

since the style will be a little different

[stoic stag]

yea i changed to 3.5 about or a little less than halfway thro and it was... noticibly different.

still very similar, but its like it lost it's personality for lack of a better description.

to get the voice i wanted, i rly only used prompts for it in the 'style of music' box.

which was these, but it didn't follow these prompts, like, at all lmao... i liked the voice i got from a gajillion generations tho so eh...
Female voice, English, Gritty raw and deep english voice, Punk electropop, J-pop synth, Happy, Passionate, Alienation

Also thx again for helping me out with this. rly appreciate it.

[sharp kraken]

Jonathan rocks - on the suno and in the chat!

[sturdy garnet]

how can I create low notes on my song ser? Tried to add "low notes" but not work

[untold briar]

I tested clearing the voice a bit, just checked to make sure it wasn't easy fix, and yeah it's not easy - at best a little better but even then melody goes out the window. also the bass comes also from style prompt, not just the audio, btw.
https://suno.com/song/9d4830ec-93fc-49c7-8154-fa7b16b87d99
https://suno.com/song/510e72b9-062b-434c-b6d0-7f216415433f

[stoic stag]

I've fixed the whole bassy instrumental issue for the most part by basically genre switching, but yeah now the voice is completely different. you are waaaaaay closer to the original voice than I've gotten so far 😮

[untold briar]

I think I could fix it with infinite patience, but to be honest, I think you'd probably be better off spending all the time and credits on new songs, even the same lyrics, like recreating it in 3.5 to start. Better end result. I was just curious at the technical challenge, and maybe it would happen to be easy, because once in awhile you do get lucky

What I actually need Suno Devs is to specific like 6.2 or 5.4 seconds. When you are trying to get just the perfect sample of a voice, it's so annoying that like 3 or 4 words is sung in just one second. HUGE DIFFERENCE

[stoic stag]

after i run out of my 10k, I gotta skip this month, so i'll buy a new 30$ sub next month. hopefully 4.0 and a bunch of new features come out by then.

[untold briar]

In case you want to try to use them, I think these are the last clips by themselves:
https://suno.com/song/2c8bc803-f73b-4231-b567-070b690af789
https://suno.com/song/0d5f1942-06d3-4df2-99c9-432e5d3c25c9

[stoic stag]

i just hope they come out with a feature that will allow me to sample a voice from something i've already created and be able to create a new song with it.

yeah im looking thro the prompts and stuff to see how you went about it right now.

[untold briar]

I lost track of where i was in the lyrics actually, constantly changing extend times

[eternal nexus]

I actually do wish for that feature to be implemented

[stoic stag]

tbh, the vocals throughout the song aside from around the end and last few verses/choruses and stuff i rly liked, its just the instrumentals largely throughout the song i didnt rly like. wish i figured out how to change them up like how u showed me, but now i know. so once i recreate it again, its gonna sound so much better. ❤️

[untold briar]

If you're curious, I think I used your Clip 1 lyrics in a throwaway segment. The reason for this is that the lyrics also effect the music and voice you get. So, if by chance I wanted to do a full clear of previous song, I wanted best chance to recreate the Clip 1 voice in its place. So even though I planned on Extending from before all those lyrics, I stuck them in prompt.

You can test this by, for example, writing a Sea Shanty song, and using a blank style prompt. Sometimes you still get a nice Sea Shanty music style.

[untold briar]

You've been checking Suno twitter fight? On the way!

[eternal nexus]

I haven't checked it but IT'S GONNA HAPPEN?!

[stoic stag]

if ur curious from a tech standpoint and want previous clips from this, i can give them to u if u wanna experiment urself or something.

[eternal nexus]

I can take my generated favorite songs voices and place them into newer songs which is quite ool

[stoic stag]

thats gonna be 10/10. my first song i completed in suno had such a unique voice i cld never replicate even after like 2k credits of trying using the same prompts.

[untold briar]

Well it can be tricky, but you can do it. I am allowed to share some test clips, we can try right now if you like, as I am still procrastinating...

Actually I didn't realize the time, but if you catch me later, I will try it if not now.

[stoic stag]

https://suno.com/song/c4be8ceb-1856-4b5e-874a-ab1e615c7206 this is basically the voice i was talking about. but later yeah, if u can figure out how to replicate this, i'm all ears whenever.

wrote this well over a decade ago O.o

[untold briar]

Try to find a similar voice in 3.5, I bet you can get a better version of that one.

[summer shore]

Something is way off within the past few days. Tonight I literally have things like "song style" "american old school rap" and lyric prompts like [American USA rapper] lyric lyric lyric ... and the outcome is a British accent 😂

I'm almost at the point where I think it's a "me problem" and I'll take a few months off and see where this goes. Nobody else seems to think there's an issue, so maybe I'm just too invested in thinking this was (previously) matching my way of working.

[topaz fossil]

i wish we could pick a voice

[summer shore]

Honestly the only way I can make sense of whatever changed is that the model is being optimized for cool things like “tap a beat” or “sing a phrase” and it will continue from that alone. But as far as text-entry creation goes, there’s been a big change for me and how I work and was using this platform.

Whichever direction this is going could still be cool and fun ... but for now I'm just not able to do what I want to do and was having so much fun with only 1-2 weeks ago.

[topaz fossil]

i mean it is better its just that i wish the instrumentals didnt change

[summer shore]

Yeah, I think for me personally I just need to study whatever prompts are working now and adapt... and take a time out to fix whatever expectations I have of what worked previously. I mean I was getting some magnificent songs using simple English. Now my typed prompts don't work and I am seeing people being commended (cool? not sure yet?) for learning how to type stuff like "∙∙·▫▫ᵒᴼᵒ▫ₒₒ▫ᵒᴼᵒ▫ₒₒ▫ᵒᴼᵒ мʸ𝓧丅"

[timid pier]

I have an issue despite setting instructions for a female voice. Both in the lyrics and the music style instructions, I still receive many versions with a male voice. Secondly, is it possible to give a command that ensures a consistent female voice style? I want the AI to use the same female vocal style for all the songs I write.

[summer shore]

It just progressed rapidly from common sense to "learn how to translate strange character sets and embed music styles into [XXX] prompts"

[untold briar]

Actually the simplest thing is Extend from a few seconds of voice with a new song. here i copied your lyics but shows it carries over in 3.5 ok https://suno.com/song/cbc0eeed-b0b9-429c-b3a4-aac8142694c0

[stoic stag]

It's better than anything i got trying to replicate it, but even in 3.5, that original voice i got is still just so... hard to explain, but to me its just so gr8. although this 3.5 voice is rly good as well. if i spent the time on this voice, it wld prly get rly good since i make tons of generations per verse/even specific lines until i get the vocals to say things a certain way i like.
once suno releases the ability to cut out previous sections of the song like udio does, it will help with this kind of stuff so much more.

that original voice i got is just so... so good...

[untold briar]

Funny thing, I rarely try to rescue or fix my own songs, I think it's better instead to spend that time generating new songs (or at least new attempts at the same song, exploring styles) - both more fun and end up with better song quality at the end of it. And a lot of the fun for me is discovering what a new prompt will do.

BUT it is also very interesting puzzle to figure out how to manipulate things and fix a voice or whatever, so I still find myself doing it all the time on other people's songs, lol

[stoic stag]

my problem is i spent so much time of my life and emotional investment into writing these lyrics, i feel like i need the vocals at least to be as perfect as i possibly can. so like with 'if i had a song' the way a bunch of the words are said like they have more personality or a flair at the end, like i make tons of generations and try to find the needle in the haystack... lels... which is kinda why its taken me about 2 months on 'if i had a song' alone.

perfectionist mentality is scary

[untold briar]

The key is to completely change up style prompts (and make [style stuff] in lyrics) to make sure you get a large range of different voices and music each time. Don't just retry, change it constantly. If you do this after spending a ton of credits you will almost certainly find a few more gems

Also use search and copy style ideas

[stoic stag]

like in 'if i had a song' at about the 4:50 mark, it took me so many changing of prompts and generations until i could get that little ending on the word gods. i tried changing style prompts, but it never changed anything at all. so i feel like if i just continued from a single clip, then changed the style like i am now, it would have worked.

so many credits/time wasted and getting instrumentals i wasnt happy with when i could have just done all the things u showed me lmao...

[untold briar]

Yeah and later I'm sure better control than that of times in Suno, so you can set the range exactly, no matter where the clips happen to start.

[stoic stag]

is there any prompt to more herd the voice into getting those little 'for lack of a better word' flair at the end of words? one of the tricks i learned is just spam !!!!!!!!!!!! lels...

although !!!! most of the time just makes the sound elongated, so it doesn't rly work... so i didnt rly use it at all until i wanted that elongation.

[untold briar]

It's not predictable but any spelling can change things.
for-lack-of-a better-word
for lack of a better WORD
for lack of a bet-ter *word*
for lack of a better... word! or put last word on new line,

for lack of a better... 
word!

just random examples.
_word_ is a good one sometimes

also theatrical styles, dramatic, musical, epic, cinematic maybe

[stoic stag]

the vocals really didnt like some words like celestial. for mythological i broke it up and it worked for most of the generations, but then it got to opposeable and indomitable and the vocals rly, rly, rly, rly hated those. that whole chorus was just... it refused to work and i finally got what i got which was at least passable. IN-DOM-IT-ABLE!!!!!!!!!! worked with all caps.

this really had a lot of vocal problems but i ended up getting it to work with this
I'll BLAZE trails across the stars to alight every PLANET no matter how far it is from a stars LIGHT!!!!!
IGNITING DREAMS within all who dare ASPIRE, my song will be the BEACON in even the darkest NIGHT!!!!!

[untold briar]

oh and dashes even in words where it is only one syllable like "wo-rd"

Too many !!! might hallucinate a bit, but if it works, it works

[stoic stag]

myth-oh-logical is what i used and it sounded mostly fine.

yeah, i noticed there was a difference sometimes when i put 5 ! and 10 !

why? idk...

like with the whole blaze one, it just refused to actually either annunciate them or it got so quiet u cldnt really hear them or it just garbled them. capitalizing them fixed it for the most part cuz reasons...

[untold briar]

I don't know music well enough, but is there a type of music formatting that tends to also specify which syllable is stressed? Like maybe opera lyric sheets or something? Also you could try accents on letters. For languges like Russian, they usually don't specify accents (I think they use capital letters) but in Suno the accents work better if I remember like стои́т

don't forget about line spacing.

For the.
Lack of.
A better.
Word!

can be quite different

[stoic stag]

hmmmm if suno is capable of comprehending special characters and accents like that, it would actually open a lot of possibilities since those characters produce certain sounds that are hard to explain with just plain romanized text.

[untold briar]

Even writing words in a different language, that sound the same. may not do what you want, but it will change things very likely

[stoic stag]

do u find stacking prompts works? like
[Verse]
[Prompt]
[Prompt]

or like
[Hopeful, prompt, prompt, prompt Verse]
[Prompt]

[molten geyser]

Hello everyone! Question:

Can Suno imitate falsetto on the male singer?

[molten geyser]

Question to avoid making 1000 attempts on something that may not work ദ്ദി ༎ຶ‿༎ຶ )

[stoic stag]

thats kinda what were just talking about. kinda... lels.

[molten geyser]

I didn't scroll uppp

[stoic stag]

basically just how to type words differently with either capitalization, -, ! or just special characters to get the voice to do something different/specific.

[molten geyser]

That may work, but I've been thinking of generating a cozy male voice that fully accompanies the song, or alternatively add a specific tag

[untold briar]

I'm not sure, but as a general tip, if Suno ignores a word, try repeating falsetto, falsetto, falsetto or even more. (Too many will hallucinate though).
Also maybe worth relinking:

I used a 3.0 prompt in 3.5 and I'm getting something more generic than the specific genre I asked for. What can I do?
#💬┃general-chat message

[molten geyser]

I think capitalization relates more to vocal improvisation with a generated voice, not the generation of the voice itself

[untold briar]

Suno help as in generally how to make songs with Suno or maybe some music topics, that kind of technical.

Not possible yet, but "inpainting" is a future feature that could do things like that, make edits or generate an intro last.

That said, if you handy with audio tools, it is possible to fake your way into something like a song intro. Basically try generating the same verse that is the verse verse of the song, immediately after the first verse (using the time feature) so it just repeats the verse, but with [instrumental] or [hook] in between. Then in audio editor, replace original first verse, sometimes it lines up close enough.

Continue has been renamed to Extend, yeah, same thing. Extend From also allows choosing custom time.

Are you seeing the word Continue in this Discord, or is it still used on website somewhere?

[sour mural]

I think I was watching a YouTube vid, but it may have been an older vid...

[untold briar]

I was wondering if there was a old help page linked by accident, but that makes sense, thanks

[untold briar]

Oh, looks like you delete question, but I was typing a intro tip. If you want to be sure you have a song intro, one way is to generate an instrumental song first, with no lyrics. Then Extend From that song after a small section of the instrumental, and use your real lyrics from that point forward.

This can be necessary when the song just always seems to start way too fast

[sharp kraken]

Well the golden rule for me has been to punctuate every phrase. So like [Verse 1. ] or like [Verse 1] . - you just dont wanna punctuate next to a bracket - kinda confuses it, so just add a space. Especially when you go from bracket to vocal n vice versa. make sure the last line of the vocals has a period, and punctuate the bracket commands - it really helps when you prompt stuff and get opossites its because suno isnt seperating out commands right. It's helped a few other people who tried this and if you have any q? just shoot me a message and i can help you anytime

4 the reccord too noneof the sigil jazz is neccessary at all for generating great stuff with suno - i just did an entire 30 minute mixtape nothing but 3.5 instrumentals

[tribal marlin]

yea definitely! i've had falsettos unintentionally quite a bit!!

for example:
https://suno.com/song/a7247c67-0b22-4a70-ae25-41c8fc6726ec
@1:15: Falsetto here
@2:20: Some really nice falsettos from here onwards, they come in and out

(the song was actually supposed to be end at 2:01, but i just let it continue since i liked some of the vocals 😆 )

[wintry saddle]

dear technical people, I was trying to help the Brazilian community and while a few rare V3 samples do exist (capybara, style Brazilian funk) most the times Brazilian in the style takes it towards samba... And funk is a scattergun of multiple genres with that name (all truly very funky in their own right!), I tried Brazilian-Funk (with and without the -) and some of the Portuguese words / subgenres of Brazilian Funk... in genre/ style, but it seems to be missing the mark.... The question from the Brazilian community, does the training data have Brazilian Funk In it, if so what's it being tagged as / how can we best summon those vibes.

Its a fun one for suno, as the syncopation and rhythms are distinct from Euro/US funk

[tribal marlin]

what are some Brazilian funk bands out of curiosity? 🤔

[sharp kraken]

Universo Parallelo

[wintry saddle]

Ludmilla, MC Poze, Kondzilla, kevinha

https://youtu.be/xBUY3VD0LqU?si=VbWC9gnvkLfNfcjQ

Very modern Brazilian funk.. but o think that's what people were going for.

[nimble citrus]

I discovered, if you put style: "Favela-Baile", play Brazilian Funk. Try there 👍

[molten geyser]

Thanks for the example 💜

Left a like 🍵

[nimble citrus]

Example, song about Cat using: "Favela" 🇧🇷 ..

[wintry saddle]

Maybe that's the route to go, avoid the word Funk

Counter Intuitive, but fine if it's consistent

[nimble citrus]

It's ok, now try find Axé.. https://www.youtube.com/watch?v=3xvGGXhMzjA

[wintry saddle]

aha 😄

[placid island]

the fuck, i'm listening this right now.....how the fuck its sounds like that

[pure bay]

Can a user extend a song made by another user?

[fiery bobcat]

how does make the vocals sound deeper

[unique stone]

Hmm. I just had "boss-fight" in my style rejected by the moderation filter due to "bossfight" being an artist name.

Which makes me wonder: Do the hyphens just get removed during parsing? I know "happy hardcore" gives different results to "happy-hardcore" but is the latter basically becoming 'happyhardcore" internally? Not that it matters much, just wondering if we can omit the hyphens... I'll try some more later

Brief testing with a two word instrument: Hyphen does make a difference than simply joining the two words

[broken lotus]

I hear a big difference when using hyphens, vs no spaces, vs spaces....

I'm not able to define that difference in words... but it's definitely there...

(wonders what the effect is on v3)

[charred merlin]

https://platform.openai.com/tokenizer

Using OpenAI’s tokenizer as a means of visualizing how Suno.ai interprets prompts;

’Happy Hardcore’ is 2 tokens, and would be interpreted as 2 entirely separate words.

‘HappyHardcore’ gets broken into 3 tokens, and this introduces a level of ambiguity into the AI’s interpretation of the prompt.

‘Happy-Hardcore’ is 4 tokens, and would likely contribute to the highest level of ambiguity, unless hyphenation makes prompts interpret differently (which would make sense seeing as how common hyphenating genres is)

[unique stone]

I used to try "happy hardcore" quite often and would always get guitar music, so I gave up until someone mentioned the hyphens, and also the fact that you can give moods or other descriptions before parts of the style prompt. So I guess this became "hardcore, but make it happy" or something.

[charred merlin]

Yeah, “happy hardcore” means something specific to you, but look at how it’s tokenized— 2 distinct words.

I haven’t experimented with prompt ordering too much, but I’d imagine it’s the same as saying “hardcore, happy”.

Adding a hyphen or spelling it as one word forces it to interpret the words together in a different way based on how it’s tokenized (readable by AI)

Curious to see what kind of differences “math rock” and “math-rock” would give…

Quick test— hyphenated definitely understood what the homework assignment was.

[broken lotus]

The hyphens have been a thing since at least v1... I discovered synth-pop was accurate and synthpop was nothing...
Then this flipped in v2, synthpop worked and synth-pop was nothing...

There's no way to write a guide for this, except to keep saying 'try whatever'... lol

[charred merlin]

I think that’s the biggest takeaway ❤️

Chuckling at the thought of you handling the v1 > v2 swap though; just grumbling, tearing post-it notes off your wall & taking the pushpins and red twine linking them all together down, like a retired detective whose just finally given up.

[unique stone]

Future versions of the wiki are gonna be like...
Q: "How do I..."
A: 🤷‍♂️🤷🤷‍♀️

[high magnet]

Hi all. Question, suno now seems to ignore all [fade out] [outtro] [end] etc... tags.... what am i missing?

[charred merlin]

Remove all lyrics and style prompts from your song, list “fade out” as the only style prompt.

Try moving the extension point to just before a verse “concludes musically”.

[sharp kraken]

Ive had luck using this - 𝇟 it is the Kievan Musical Annotation for the end

[fast scaffold]

It would be nice if we could have a referral scheme. I've told about 10 of my mates who now use Sona daily

[charred merlin]

Is there a trick to using the 𝇟 symbol?

I've tried a few examples; v3.5, a single 8-line verse, followed by the symbol. In each attempt, the song would have reached the 𝇟 symbol around ~0:40, but the songs continued to generate to 3:07 & 4:00.

As a humorous side note, I've found that v3 has a higher success rate at wrapping songs up rather than extending the denouement into a whole second act, but I loved this little tantrum it threw before deciding to end

[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

[torn pollen]

I've been trying out the 3.5 model and there seems to be a bug in creating the bridge part: in the 3.0 model, verse 1, verse 2, and the bridge part are created only once each, but in the 3.5 model, the bridge part is created once more with the lyrics of verse 2 before the bridge part I originally created. I think it should be fixed.

[pure bay]

Does anyone know?

[broken lotus]

yes, if that song is set to PUBLIC.

[pure bay]

Can one just post the link to another user, instead of setting it to public, to make it extended by that user?

[broken lotus]

No.

And it's against the ToS to have more than 1 free account.

[pure bay]

I mean 2 different real people.

[charred merlin]

https://suno.com/song/c1adb531-1261-4e11-a734-2a12161a1a0b are you able to extend this?

(still generating, but it's private with the link shared)

[broken lotus]

yes, collaborations... we still have to 'unlock' the song by making it public...

it is not ideal, I think, I see a difference in 'share to listen' and 'share to collaborate'

[broken lotus]

yes

[charred merlin]

share to listen vs share to collaborate
That difference is why I'm not participating in Summer of Suno, and keeping my songs private to distribute. I would LOVE to see that feature added before the end of June so I can participate, but I know that's a tall order in such a short amount of time.

[pure bay]

please read my private message.

[charred merlin]

https://tenor.com/view/sweating-nervous-wreck-gif-1107534431268855805

[wheat rose]

asked in General but it got lost in the noise so i'll ask again here.. Question for any youtube publishers... Are you marking your music videos with their "Altered Content" flag for content made by AI? it seems to me that they intend for it to mean something entirely different but the wording they used clearly would apply

[charred merlin]

I'd say that differs from Instagram's "Label as Made with AI", where they're probably wanting people to admit their content is a deepfake upfront rather than be responsible for backtracking.

[hot fossil]

Can we please get the bot trained on Jeremy Soule😭😭😭

[wheat rose]

According to their Help Center This could include content that is fully or partially altered or created using audio, video or image creation or editing tools.

so my reading is that literally any content created or edited by software would need to be marked, a plain reading is even if you just use a video editor or audio normalization

[charred merlin]

Can you link to that page? I'd still imagine the context is for "altered" content that could be reasonably construed as being published by an existing artist or person.

Though it could just be an underdescriptive title/overdescriptive definition

[wheat rose]

https://support.google.com/youtube/answer/14328491?hl=en

[charred merlin]

Well, there ya go

[hoary bronze]

How do I make a section where the al speaks the lyrics out but not ‘singing’ it out in middle of a song

[broken lotus]

This part is about the intent.

[broken lotus]

[spoken]

[charred merlin]

While the majority of the content on the page seems to indicate they're more worried about misrepresenting real people, but there absolutely is a statement that we need to disclose synthetically generated music.

[hoary bronze]

Is that the same as [Spoken Word]

[wheat rose]

yep there is it just had to scroll down farther 🙂

[broken lotus]

yeah YT can bite me, tbh.

[wheat rose]

I have been disclosing it in all of my videos since the plain reading implied I needed to already

[charred merlin]

"spoken word" can imply a jazz-feel, like William Shatner. "Spoken" just indicates not sung

[hoary bronze]

alr

[broken lotus]

so... 'reverb'

[charred merlin]

You've better not've been adding reverb to your tracks and making us believe that you were able to do that in realtime 🤬

SCANDAL

[broken lotus]

oh wait... that is under DO NOT NEED TO DISCLOSE...

(not trying to misrepresent, I just didn't see it)

[wheat rose]

my stance is generally that the rules are stupid but easy enough to follow and not worth breaking 🙂

[broken lotus]

My stance is that my creative process is none of their damn business and they don't pay me for ads they serve against my videos.

[wheat rose]

My intention is that they will eventually be paying me for them

I have other monetized youtube channels and i'm not willing to take needless risks 🙂

[charred merlin]

Yeah, if you're already in the ecosystem best not to make ripples.

[hoary bronze]

I tried
[adventure intro]
[Emotional Spoken]
“My lyrics”

But the background music is still very loud and I can barely hear the spoken lyrics

[charred merlin]

That result makes sense given what you asked it to do, especially if the intro is loud and adventurous, then you're expecting everything to go silent to accomodate the vocals without having communicated that.

AI works best when used as a collaborative tool alongside human intelligence and creativity.

Might need to create the song in multiple parts; 1) the adventurous intro, 2) the fadeout to a soft background melody, 3) the spoken content

While normally I suggest extending from "Get Whole Song" to preserve cohesion between earlier segments, I'd consider just extending from extensions since each part is thematically different from one another.

[eternal plaza]

Is there to download and run the Suno AI music model locally?

[charred merlin]

😛 No

Bark, yes

[sonic timber]

can someone of the support team help me with something serious/

[naive quail]

It'd be helpful to post what the actual problem is.

[unique stone]

Please don't post the same thing across multiple channels. I see this in the general channel, off-topic and showcase as well as this technical discussion channel.

[charred merlin]

[e-mail]
[e-mail]

[sonic timber]

bro it's literally a question i aint senting an email lol

[neat nexus]

@rain shuttle Thanks.

Once was good.

Did you leave it up in one channel at least so we can check it out?

You did. Nice work!

[broken lotus]

you can DM me if it is private.

[soft gyro]

is there any meta tag to tell the AI not to use drums in a song for exemple?

[solar rose]

If I change the cover image, is it normal that the original image is still used for the video file and not the new one?

[unique stone]

Yes, the video gets rendered when the song is created and doesn't get updated when you change the image.

[neat lake]

Makes sense, thanks for being civilised about it 🙂

[solar rose]

That's a pity. If I change the cover image on Udio, the video is re-rendered before downloading.

[unique stone]

Maybe it'll change in future. It'd be nice, especially if we get to edit lyrics after generating the song and have that update the video too, because that can look a mess and not follow the actual song sometimes 😄

[snow pelican]

/chirip

[neat nexus]

Chirp is retired. Use the new model at https://suno.com/create

[agile veldt]

I've noticed something weird about the videos that Suno generate. The audio appears to be 192 kbit/sec AAC, but they sound worse than the 192 kbit/sec MP3 files that you can download. I would have expected AAC (mp4) to sound better than mp3 at the same bitrate. So that's pretty odd. 🤔 Here's an example of two AAC (mp4) files. The first is encoded by Suno, the other one is encoded by me. They are the same bit rate and same format, and thus the same file size.

For instance, notice how the crash cymbal at the start "flickers" in Suno's version:

Suno's aac encoder is like "frfrhrhrhh frhrhfrhtr hfrhrhthhrrr"

All the transients (drums) have a weird blurry treble. At least if you listen on good speakers.

I just used ffmpeg's default settings -acodec aac -ab 192000

[rain shuttle]

is there any advanced prompts i can play with

[unique stone]

This was something I noticed fairly early on when v2 was the only thing available. I stopped downloading the videos because of that. It did improve. Did you encode from the WAV? I wonder if it's generating the audio for the video from the MP3. One of the devs did suggest before that the video had higher quality audio but I've not experienced that.

[agile veldt]

yeah i encoded from .wav

i also downloaded the mp3, which was almost as good as the .wav

[rain shuttle]

cause the genres are not matching up

[agile veldt]

not quite of course

but the mp3 is closer to the wav than it is to the video

maybe the video's input is another lossy format. or maybe the encoder is broken.

(never re-encode a lossy format to another lossy format!)

[rain shuttle]

deep house sounds like electro pop

and all the edm stuff now is like cotton candy pop

[agile veldt]

there's probably a lot of pop in the data set, which makes pop show up a lot. and to make things worse, i think suno has been using people's votes to change weights in the model. and if you go by popularity you get pop (hence the name), so that's bad idea IMHO.

[crystal prairie]

cc @regal vale

[agile veldt]

@rain shuttle try "custom mode" + "instrumental" + this prompt:
1998 tech-house, deep house

i just tried this with v3.5 and got pretty decent stuff. now i'm gonna try v3.0 instead and see...

[unique stone]

The hyphens strike again 😄

[agile veldt]

i never know whether they're a good idea or not! 😄

so i did one with, and one without

v3.0 and v3.5 seem fairly equal when it comes to this little test

[unique stone]

As was being discussed earlier, it's kinda like:
"happy-hardcore" = happy hardcore genre
"happy hardcore" = hardcore genre, but happy

[agile veldt]

v2.0 gives me slightly more odd sounding results as expected from an older model

@unique stone ahhhhh now that makes a lot of sense

i guess the meta data must have had hyphens then

[unique stone]

Sometimes you can prefix an instrument or genre with a mood/feeling. I throw things like "mysterious bass, curious trumpet" at it to see what happens sometimes

[agile veldt]

For more down-toned techno'ish house, this is giving me good results: "1998 deep-house, dub-techno" (custom mode, instrumental, v3.5)

[unique stone]

Sometime I might try "sad happy-hardcore" 😄

[agile veldt]

Genre generator: https://www.robotplanet.dk/humor/word_generator/?id=2

[unique stone]

Gonna bookmark that as I get stuck trying the same handful of genres often

[agile veldt]

i wonder if i should change "hip hop" to "hip-hop" or "hiphop"

[unique stone]

I usually go for hip-hop when I mess with that

[agile veldt]

okay, hang on...

fixed.

Ooh, good prompt: "2000s deep-house, dub-techno, cumbia trival"

[rain shuttle]

so this is what i got with that prompt

[agile veldt]

sounds about right... 909 hihat there... but vocals!?!

[rain shuttle]

vocals come in later

[agile veldt]

And if anyone aren't certain what a real 909 hihat sounds like, here's the real thing from the actual machine:

[rain shuttle]

i used 3.0 3.5 and instrumental and custom 2.0

if there is anything i can do too help the model let me know i am very passionate about music and this project

it just trails out to a static sound

[agile veldt]

maybe whisper sweet little nothings into its ear

no, i have no idea 😄

[rain shuttle]

hell yeah i will do that .....lol

I will tell u 1 thing it has given my poetry a voice thanks

who do i talk too about donating to suno i want to donate my record sales and earnings from my music to suno. i can explain why in private but who do i contact?

if someone can get back to me about this i would be grateful

[worthy monolith]

i have a question

https://suno.com/song/9886de2f-4148-4605-beaf-c69e55cd3f3e im wondering wether it is possible to tell suno ai to sing certain lirics in certain ways: for example singing the chourus at the and of the song the same way as in the beginning of the song along with the same or similar instrumental rhythm to it

[agile veldt]

putting [Chorus] in front of it helps it to recognize that it's a chorus

https://www.suno.wiki/faq/metatags/verse-and-chorus/

[worthy monolith]

i've been doing that im asking what can i do more?

[agile veldt]

that's the only trick i know

[worthy monolith]

ok

[agile veldt]

maybe be careful that the meter in your lyrics fit the rhythm

if it does, it helps the AI to fit things into place

[worthy monolith]

meter?

[agile veldt]

like, the amount of syllables and words

the "rhythm" of the text

[worthy monolith]

i dont have a singers mind so thats gonna be difficult for me to figure out

most of my lyrics are generated

[agile veldt]

yeah that's gonna work poorly

you need to have a singer's mind to make it work really well

chat gpt doesn't have that

[sharp kraken]

I was lucky enough to start making music on a 909 in 99 lol

[floral prism]

Hello all. Any advice on ending a song? I'm simply burning through credits extending songs in an attempt to get a simple fade out. I've tried every tag within the lyrics I can think of, outro, ending, fadeout, fade to silence. I've use these in an instrumental style prompt too, and I get zero results. what can i possibly do different?

[agile veldt]

End a song? That's a common problem people have... it's this common:

[sharp kraken]

𝇟

[agile veldt]

https://cdn.discordapp.com/attachments/1235188700728787006/1248196232090746930/stop.jpg?ex=66641a54&is=6662c8d4&hm=3da6dcd698d32fb5065de28db35450bfda6777ce26f9745261c598e6455a6fbf&

[sharp kraken]

Put tis at the end of your lyrics 𝇟

[agile veldt]

[End] and/or [Outro]

[sharp kraken]

that will cause like imeadiate cut

[raven turtle]

Try [Chorus repeats]

[sharp kraken]

so make sure you manually fade the instruments

[floral prism]

no dice on this. I've tried it multiple times.

[hybrid linden]

guys if you want to make an extended mix of a house music track, hos do you make suno create an intro and an outro for the track?

[agile veldt]

whoa ... 𝇟 ... what sort of non-ascii character is that. hang on...

[worthy monolith]

i'll try it

[floral prism]

will attempt this

[sharp kraken]

Thats the music notation for Kiev

all stop

[agile veldt]

that 𝇟 character is so obscure that my text editor can't even display it 😂

[floral prism]

how do I even type that without copy pasting?

[agile veldt]

you probably can't

[sharp kraken]

look upalt codes for music notation on google

[floral prism]

unfortunately it didn't even work for me, so that's a no go too

[agile veldt]

leave it to python devs to invent languages that require whitespace and non-ascii as a part of the syntax! 🙃

[sharp kraken]

you can you hold alt and punch in a 6 digit number

[tacit thorn]

Any idea on how to tag specific lines, verses, or words to get different voices consistently?

[agile veldt]

gosh, i found it: https://www.compart.com/en/unicode/U+1D1DF

[sharp kraken]

ensure you have proper punctuation

[agile veldt]

CURSED

[tacit thorn]

I do

I made 4 generations, with a verse that had a call/response tagged as male and female line by line

1 did it, 1 switched every two lines instead of every 1, and 1 read the description word as if it was a lyric even though it was bracketed

[sharp kraken]

it can be tough. So you have to have everything your using in the style prompt

okay you have a punctuation error there

[tacit thorn]

I ended up having the iteration I wanted but had more junk/error/problematic generations than usual

I'm trying to figure out limits and possibilities

Here, I'll copy/paste what I was using

[sharp kraken]

make sure every phrase has either comma or period

[tacit thorn]

One was:

[Hook, male]
Who's the biggest baller in all the land?
[Female voice]
It's grebby grebby G baby he's the man
[male voice]
I said who's the biggest baller in all the land?
[Female voice]
It's grebby grebby G baby he's the man

[sharp kraken]

[verse 1. ]

or [verse 1] .

[tacit thorn]

Another: Gonna get knocked out get your ass to the bank
Hey T pain, this loser can
[autotune, slow]
buy us a drank

[sharp kraken]

punctuate every single phrase and i bet it will work

[tacit thorn]

Cool, I'll give it a try

The results were odd enough that I wasn't sure how to cure

But I appreciat ethat input

[sharp kraken]

[autotune, slow, ] "lyrics" . (thats how i use those

[tacit thorn]

I was able to cure some of the misplaced break/bridge stuff that had been popping up in songs with instrumental breaks

[sharp kraken]

yeah as long as theres a period between the lyrics and the []s you shouldnt get run over like that

[tacit thorn]

Ok awesome

Thanks Josh

Why can't you send them here

Can send here and then delete them after getting feedback

[left geode]

why almost all my songs sound like old tapes from 60s?
I tried different devices but the quality is so bad, can't even listen it, it hurts. 😦
voices sound non human, instruments are distorted. There is like a hissing sound especially during intense moments.
Typically it starts Ok, but degrades towards the end.
I am a new customer. Is this normal and it generally can not produce anything good or I am doing something wrong?

[tacit thorn]

What kinds of tags are you using

I don't have that experience at all

[supple dagger]

guys how do i generate music can someone tell me plz and thank you

[unique stone]

The quality will vary, but there is a limit to how good it will sound. In future I expect this will improve. There can be degradation as the song gets longer - the only way around this at the moment is to extend from a time before this degradation starts.

[weary wharf]

Long songs always suffer. As long as the start quality is crisp, at least it will continue up to 2-3 minutes.
Depending on the instrument count, gain on the volume, the noise is generated and bleeds into the upcoming sound.

With limited instruments, "low volume" generations it is possible to keep the quality alive.

Else if it's not degraded, you can use tools like RX 11, Ozone or some other similar tools to correct the noise and recover some of the lost harmonics.

[tardy hatch]

Has anyone found any specific tips on getting a better country singing voice? I find a lot of times that I will get a smooth more realistic voice on other genres, but as soon as I do a country song it very obviously sounds like a robot cowboy and doesnt really do a country accent or sound very well. Wasn't sure if this is just the way it is until AI improves a bit more, or if anyone has any music style tags they use that they have found gives them a better country voice?

Actually, I have just noticed in general that I can't get any country style to really sound the way I would like it to. I am a fan of like 90s country and I can't seem to get that same feel

[naive quail]

Maybe try some variations like Early Country or Honky Tonk?

Bluegrass gets some of that southern drawl

https://suno.com/song/a840f451-bb9a-477b-946f-f5b8e7889a1d

[warm coral]

generally quality is somewhat lacking at this point but it is getting better and better. it's kinda like in it's first year midjourney phase

[tardy hatch]

Still couldn't get it to do what I wanted to, so I just tried a reggae and island mix and was pleasantly surprised.

https://suno.com/song/cbdeeda5-96e2-4f67-a1e6-b78ad4a6b83c

[naive quail]

Maybe not THAT bad. At least when I have "hands" in the lyric's prompt, the singer doesn't sing "FREAKISH TENTACLES!"

[neat nexus]

southern accent?

[rain condor]

I am also experiencing this. It seems to be more noticeable on quieter more dynamic songs. The audio is “fuzzy” sounding. I’m assuming it’s aliasing from the data compression that’s being used. Unfortunately it’s not fully usable for the stuff I’m doing.

[dapper sphinx]

I have a question, can I use the music I've generated with Suno for a project (professional and where I earn money) without thanks them (sorry for my formulation, I'm french🤣 and I wasn't don't use a traductor)

In summary : Do I own the music generated or not

[charred merlin]

Did you use a free or paid account?

[dapper sphinx]

Free account

[charred merlin]

Suno owns the song.

[dapper sphinx]

Oh s**t it's not a good news for me 😭

[charred merlin]

Even if you pay for an account, they own that song.

[dapper sphinx]

Ok

[unique stone]

Yeah as soon as I found out about that I subscribed immediately, even though I don't sell my stuff

[dapper sphinx]

But if I make almost the same music myself it’s good right?

[charred merlin]

Oh yeah

You could subscribe and attempt to recreate the song, and THAT will be yours. We're not really sure what the legality behind extending from a free user's song and downloading THAT result, but I feel like that's specifically trying to avoid attributing credit.

[dapper sphinx]

Ok thanks 👍

[charred merlin]

Sorry for rambling, but I just recalled another detail that might be relevant to you-- you retain ownership of any content you put into Suno, so you still own the lyrics that you used for the original song

[dapper sphinx]

Oh it's fine but I've don't use content to generate 🤣

Uniquely explication of what I want 😅

[sharp kraken]

no one can own what is generated only through promting. - so i could grab anyones Suno track and do whatever I want with it, and it would be really hard to control it.

[hidden hare]

hello, not sure whom to ask and where to ask but i just subscribed and I dont see the fonts are visible unless i highlight it. is there any workaround?

under the create section while trying to write my lyrics

any help will be appreciated

[sharp kraken]

to save yourself a lot of headaches i suggest you write prompts into a txt editor

[unique stone]

With a paid subscription, Suno gives you all the rights they would otherwise have over the content generated from the prompts you input (at least, that's kind of how the ToS phrases it.) They allow you to use the content generated by the service in the ways described, so for example if you monetize a song generated on the free plan then you're breaking the ToS.

The generated audio has a watermark on it as well so Suno may be able to determine whether a song was created under a free account or not, and potentially take action if they wanted to in such cases.

Copyright is a different matter.

[broken lotus]

You would have a hard time distributing it, when it's already protected by a distributor.

[sharp kraken]

Yeah I know I publish all my tracks as is on youtube so i can have some kinda control - but i feel like if you really want to own them you gotta take it to post

I know I as a very small peanuts creater would have a helluva time stoping someone from using my shit, and it's probably all over spotify already

[broken lotus]

By the time it went to court for copyright... 😏

[sharp kraken]

it wouldnt take much just to bury someone in my position

im just hoping i can hang in long enough to montitize something at this point or get a remote jobas prompt enginneer

[charred merlin]

I would love to be wrong, but I don't think "prompt engineer" is a real job title outside of ChatGPT-based self improvement tutorials that people charge for.

Even then, I would imagine it would involve A/B testing instead of experimenting through ascii art.

Maybe you could sign up for one of the paid feedback interviews the next time they run then, and use that $20 to sign up for the Pro version of Suno and you wouldn't need to worry about your content hanging in for long enough, because you'd own it! 🙂

[soft gyro]

ha ha, it happened to me too today also 3 times.... I wrote [end] but it took 2 more minutes to finish

[soft gyro]

it happens to me also to certain music styles that the quality of the sound get worse and worse

[sharp kraken]

Dude stop it - i dont want to talk to you so quit talking about me

quit making memes about me

your bullying and harassing me

[dusty stag]

Hi all, just discovered suno and I'm in love with it. Was wondering if there is documentation somewhere with specific formats we can pass on the prompts to give instructions? As can we fine tune melodies in songs etc

[tribal marlin]

yea, i initially had some examples posted with favela, baile and carioca, but because i'm less familar with the style i deleted them in case I had the wrong sound 😆

removing words is essential in these situations since the word "funk" itself will might have too strong associations. So if you include the word "funk" it might have too much of an impact and pull away from the favela, baile etc. That said you can stack the tags up, and put terms like "funk" right at the very end. It seems to have less weight when placed further in the list

[sharp kraken]

♂🎙, ♀🎙 these work for male/female vocals and take much less space than words (Bonus Tipsss)

[tribal marlin]

haha really lol???

[sharp kraken]

yep

🎷🎻🎸🎺🥁

theese too

[tribal marlin]

lol no way haha, intriguing

[sharp kraken]

it saves you a lot of letters

but emoji is 2 chracters

[tribal marlin]

should just do some songs that are only made from emojis

[sharp kraken]

hehehe yeap....

you could have like 20 different vocalists

add duet, beatboxing vocal fugue

makeit a barbershopmquartet lol

[tribal marlin]

ahhh this does make sense the more i think about it. I guess the emoji's all have strong relations to the word they are associated with

[night tundra]

Can we share other music that's homemade?

[sharp kraken]

yup works for notes and all kinds of things

[tribal marlin]

that's a super cool tip actually, i am going to definitely make use of it, at the very least to experiment with having a ton of tags

[tribal marlin]

has to made with Suno, or at least somehow related to Suno

[lament cove]

Just wanna say that sad adjective in the style prompt field always make banging songs 🤣 🔥

[unique stone]

Suno tends to involve a lot of experimentation, and the results vary (there's no guaranteed way to get a particular thing to happen) but there is a wiki with some guidance https://www.suno.wiki/ which is a good starting point 🙂

[dusty stag]

ah ty! I was looking for some documentation but I could only find the faq

:)

[unique stone]

Beyond that, if you get stuck with anything in particular, ask here and someone may be able to offer some advice (or in #💬┃general-chat but that can be a bit too noisy sometimes)

[dusty stag]

awh ty. honestly I'm amazed this kind of service hasn't blown up out of propotion like how openai did lol

[tribal marlin]

it's well known... but it could be even more known 😆
Like everybody interested in AI knows about it, but the mainstream awareness hasn't fully kicked in yet (though only a matter of time)

[unique stone]

Heh... A couple of months ago there was an influx of new users and there were a few days where the service was so overwhelmed they had to temporarily make it availably only for subscribers otherwise nobody could do anything. Think they scaled things up a bit since then so it works well most of the time now.

[glacial scarab]

I member "Elevated Usage"

[unique stone]

Oh how we meme'd

[charred merlin]

Oooh, this has grown a bit since I've last checked it out 👀

[outer token]

Has there been any reliable way discovered to actually end a song instead it extending on a minute plus after you tell it to. end, fade to end, fade, stop, Or whatever that little symbol was posted yesterday that was meant to stop it, 10 seconds.. They all seem rather miss at the moment

[hollow hare]

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

.
IIRC, someone did say you can also end with something like [10 seconds] in 3.5? Have to look it up again: #📚┃suno-school message

[sharp kraken]

END𝇟

thats musical notation'

works pretty goood like 4/5 end

[outer token]

I will give that a try thank you. I've tried the 10 seconds It only worked like twice out of twelve tries

[outer token]

Yeah, you are who I saw suggested yesterday when I was in here. It never worked for me, But great to hear that it was reliable for you

[warm coral]

how to get something that sounds like radiohead?

or disonance but not unlistenable

[sharp kraken]

Yet

like Dissonant Yet Danceble

Complex Yet Smooth

etc..

[ruby condor]

Why doesn't 'Extend' work the way you would think that it should work? It just creates a new song. Thats not how extend should work. It should take existing song and........ extend it.

[naive quail]

it creates the extension as a separate track. Once you get an extension that you like, you can merge it back into the original via the menu option "Get Whole Song":

[ruby condor]

thats dumb. Why make us go through the extra steps? I don't think it was always like this. I don't like this.

[naive quail]

🤷 I prefer it this way... or I would if Suno didn't happily hallucinate its way back to old versus and chorus' so that we're forced to find a time to backtrack to.

[outer token]

Just tried it on 4 that were messing me up in the past 24 hours. That worked like a charm. Thanks a bucket load

[hollow hare]

Likely resource management. Less resources taken to merge parts you select rather than doing it twice every time you extend when, most times, it would be needless

[ornate spire]

it's also more convenient to immediately listen to the new part rather than having to seek through the whole song for it

[tough parrot]

why the bot always says does not respond? please help

[sturdy garnet]

can I ask how to create 2 different voices on 1 song?

[outer token]

[Voice 1]
Lyrics
[Voice 2]
Lyrics

That works for me more often than not

[sturdy garnet]

Thank you ser

[hollow hare]

Chirp bot is retired, use the website https://suno.com

Some helpful starting links: https://suno.wiki
https://youtu.be/lj8utBd_IUM

[golden cove]

i have a question, my songs vocals slowly get worst and worst in audio quality, when i generate a song i press like and diss like or nothing but i never delete those, my question is , it is possible that the AI is "learning" the songs i generated so far and since I never delete them, the majority are of low quality so just learned low quality ?

[hollow hare]

The issue is that the AI does not remember the original style used. What it does, for each extension, is study the style ~1 min (unsure if this got extended to 2 min for v3.5) prior to the continuation point to learn/mimic the styling you had (which is why extensions in v3.0 are 1 min long as the other min was reserved for studying).

So if you had a loudness or degradation issue within that minute, it would start becoming emphasized as that is what the AI picked up on and all future extends builds up the issue as long as it exists within the 1 min of continuing.

It also explains why the style can change when you vary the starting point by a few seconds

The down votes do help with future model training though so keep it up (it has no immediate effect AFAIK as the models are frozen in their release)

[graceful condor]

My credits don't increase daily / weekly as is stated on the subscripion page, is there a reason for this?

[hollow hare]

Subscription plans are month to month renewals

https://www.suno.wiki/faq/my-account/why-subscribe/

[graceful condor]

Ah I see, that seems different than when I started but maybe I misunderstood something. I suspect this changed in the last couple of months?

[hollow hare]

It has been like since the start of the plans as far as I know (but I have not been here that long). You can ask in #💬┃general-chat with the more active user base

[vestal rapids]

It's been like that since at least December last year

[graceful condor]

In any case that explains everything, thank you!

[golden cove]

that a really unfortunate thing, i mean its explain lot of things when i extended a song and this is help with it , but its not explain the overall problem when i generate a new track its more more rarely to get good or high quality vocals and often i get the same low quality with the same "radio" effect

[hollow hare]

RNG in many of the cases if it is there from the start though I did notice that degradation increases past the 1 and 2 min marks.

Using the style "clear [male/female] vocals" worked for me in 3.0 when I wanted clear voices. I haven't tested that with 3.5 (but it only helps so much)

Some people do small extensions at like 30s-59s to maintain quality

[hybrid geyser]

質問があるんですが日本語で話せる方いませんか？

I have a question. Is there anyone who can speak to me in Japanese?

[hollow hare]

https://discord.com/channels/1069381916492562582/1154780866736820345

[hybrid geyser]

Thank you so much!!

[golden cove]

this was my trick to in v3 XD but not work to much in v3.5 , just like in v3 i can re creat the early 2000' nu metal are but with v3.5 i cant get anything i do , but yes RNG is a really big part of this

[hollow hare]

Maybe using a hyphen instead of spaces for a compounded terms will help get the correct genre (e.g. nu-metal)? It is something JFly noted... though there is a chance that the AI may have unlearned what nu metal is in 3.5 like with phonk from what I read.

Jonathan Fly's post regarding specific genres sounding generic in 3.5: #📚┃suno-school message

[golden cove]

thats a really usefull post , thank you

[tame mango]

Need some advice on some pronounciation of language that have more vowel variations for same letter. For example, I create this song in two language Bahasa Indonesia and Sundanese. Sundanese vowels contain some variation of E sound (é like in English 'bed', e like in English 'pearl' and eu that not have English equivalent, but closely to korea ㅡ (eu, IPA ɯː). The problem is, SunoAI not consistent in it's pronounciation, sometimes it read corectly but at other part it wrong. Example from the song, 'melengking' are both sung as e 'pearl' correctly, but in later part 'lengkingan' with same root-words sung wrongly and intead using é 'bed'. Any tips to make SunoAI know the difference? For eu, using written eu is often fail, so I use IPA ɯ and it works most of times. Thank you. https://suno.com/song/3c19d806-e8fd-4ec7-842f-fd287f2d37a9

[hollow hare]

Don't forget the ultimate power move people use when 3.5 does not work properly:

• Generate a great start on 3.0 with the style you want (usually the intro instrumental)
• Make the extensions in 3.5
• Make the termination extension with 3.0 because 3.5 listened to all those songs asking the DJ to not stop the music: #📚┃suno-school message

[golden cove]

hmmmm.... I tried a couple of times start with 3.0 and extend with 3.5 but the switch is really noticeable cuz its technicly switch style, may i try this time harder cuz i wast 1000+ coins for one song and its really bugs me to not work what i want XD when we are talking i try it on a free acc with a simular song what i want and it change style so drasticly in a old nu metal style , with other style not have the same problem just like the AI forget what is nu metal is

[graceful condor]

I had to refresh the web page on Suno, I press play, I stop, I go to different points in the songs, but then it got stuck, would not play, so I would refresh; the problem might be my Firefox web browser and perhaps not Suno related but I don't really know

[tame mango]

oh, additional info by the way: sometimes IPA ɯ is read as u (English double o like in 'book') if used at beginning of words, but often times its work as intended 'eu'. So I'm very confused how the correct usage of including IPA symbol in the lyrics. Similar Korean ɯː 'eu' also give mixed results.

[molten geyser]

Hello everyone! 🌟
How can I make the tempo of a part in a song very fast-paced? Similar to the narrator stating terms of a product at the end of an advertisement

And, what tag can I use to trigger a female narrator (or news reporter) monotone voice?

[hollow hare]

Try uptempo or one of the Italian tempo phrases like prestissimo (see #📚┃suno-school message )

[hollow hare]

That should work though you may need to resort to making a new extension for it with the voice type in the styles prompt (and repeat it so it gets weighted more. E.g. female narrator, female na...)

The speech metatags in Suno I am aware of are [Announcer], [Narrator], and [Spoken Word] and their variants (e.g. female narrator).

There is also the following styles you may want to consider: standup comedy. Unsure about live audience and live performance

https://www.suno.wiki/faq/metatags/voice-tags/
Old metatag list: https://suno-ai.notion.site/Using-Metatags-34944efe09ec41d693e314f13af44695

[molten geyser]

Thank you! 💜

[mild pond]

what image generator does suno use in the backend?

[vestal rapids]

DALL-E I think

[mild pond]

I dont think thats the case, dalle3 in the api takes around 15 seconds, but suno creates the image in around 5 seconds

[vestal rapids]

I don't know then

[coral plover]

how is suno doing the small videos when posting the links here? with lyrics and all

or is that what one gets when you download as mp4?

ok got it. didnt even realize 😄

[broken lotus]

I think Suno uses SDXL... probably with a lightning or hyper lora to speed it up.

[mild pond]

yeah, thats also my thoughts,but im not sure

maybe sdxl turbo

[ocean grotto]

why is my track default length always 32 seconds? cant i make it longer somehow?

[unique stone]

You can press the "extend" button to add new parts and then join them together later by using the "get whole song" option on the latest part.

Are you making songs with lyrics? If so it might work better to put more lyrics in at first so the AI has more content to work with.

[ocean grotto]

oh okayu

so after i add intro, and extend the song, i use "get whole song"?

i cant see the "get whole song" option

[unique stone]

You can't add an intro later (intro will be the first part then you extend from that)

I just want to check, you are using https://suno.com and not something else, right?

"Get whole song" is on the "..." menu for an extension ("part 2" etc.)

[ocean grotto]

yeah, alright - lemme check

[solar rose]

I have an yearly subscription. What actually happens to the unused credits? Do they expire at the end of the month or are they carried over to the new month?

[broken lotus]

the credits do not rollover.
use them.

[graceful condor]

Can I ask about payment here?

[onyx token]

I have new lyrics for a public domain song. I uploaded the lyrics and the suno voice speaks them, but i don't know how to "teach" it the melody. Any pointers? TIA

[unique stone]

It might not be something that can be solved on discord but ask and maybe someone will have the answer 🙂

[graceful condor]

Thank you for answering.
I wish someone will answer, but I try to do myself. thank you!

[grizzled slate]

Hi, I wanted to try mastering a track using LANDR, but when I try to upload any of my Suno files, it fails. Any ideas or suggestions would be gratefully received!

[daring lava]

When experiencing issues with uploading files to LANDR for mastering, there are a few common factors to consider:

1. File Format Compatibility: Ensure that the audio files you are trying to upload are in a format supported by LANDR, such as WAV, MP3, or AIFF. If your Suno files are in an unsupported format, you may need to convert them using an audio converter. (This shouldn’t be a problem unless you formatted it differently on your end prior)

2. File Size Limitations: Check if there are any file size limitations on LANDR. If your files are too large, you might need to reduce their size by exporting them in a lower quality or splitting them into smaller segments.

3. File Corruption: Sometimes, files can become corrupted during the saving or transferring process. Try re-exporting your Suno files and see if the issue persists.

4. Internet Connection: A stable internet connection is essential for uploading files. Ensure your connection is stable and try uploading again.

5. Browser or App Issues: If you are using a web browser to upload, try a different browser or clear your current browser’s cache. If you are using a LANDR desktop app, make sure it is up to date.

6. Account Restrictions: Verify if there are any restrictions on your LANDR account, such as storage limits or subscription requirements that might prevent uploads.

7. File Naming Conventions: Ensure that your file names do not contain any special characters that might cause issues during the upload process.

If none of these steps resolve the issue, consider contacting LANDR’s customer support for further assistance. They might be able to provide more specific guidance based on their system's requirements and capabilities.

[grizzled slate]

Hi and thanks for your prompt reply. I followed all the suggestions via the landr site, but nothing worked. I did contact landr but not had a reply. I don't know how, but could it be that it is Ai generated!? Are there meta tags within the file? I used to use Logic, but those days have gone. So, I'm looking for other options.

I did try BandLab, but to be honest, I think this just turns up the volume on everything!

Just looking at Meta Tags on Wav and it does show that it is from Suno.com

[charred merlin]

Just seeing this now, but you tried using both the .wav and .mp3?

[grizzled slate]

Yes tried both.

[charred merlin]

Paid Suno account, or was the song generated for free?

[grizzled slate]

Landr... was trying before buying....

Paid Suno..

[daring lava]

Never heard of this being an issue. Perhaps removing the Meta Tags could help, maybe give that a go and see if it helps in any way. If not, please let me know

[charred merlin]

I did contact landr but not had a reply

[grizzled slate]

It seems that Landr require sample rate 44.1, but looks like suno track 48 @ 16bit. So, I converted to 44.1 @ 24 bit... but still no good. I'm giving up on landr now.

[grizzled slate]

I contacted Landr on Thursday, so maybe why not heard.

[unique stone]

I did a little testing to see what effect musical notation unicode characters may have in the lyrics box.

Things I tried include:

• Moods (happy/angry/sad emojis)
• Male/female emojis to indicate singer gender
• p/m/f notation to indicate dynamics
• increasing/decreasing dynamics
• Musical notation (note lengths, rests)
• END𝇟

I'm not convinced they have any effect. Sometimes they get "sung" as if they are a word, and if you don't put a space between them and a word it will often just skip the word entirely.

I eventually tried using less of them in case there was just too much stuff going on, and also putting some of the symbols inside the square brackets.

https://suno.com/song/13b95396-8d0a-451d-9f7e-a6faa2bbf214
https://suno.com/song/d127f9b0-d646-4593-9121-77b1bffc545b
https://suno.com/song/ab677f03-07dc-43b9-861f-06c8e4ccf760
https://suno.com/song/866335ae-0fc2-428d-92d3-92072ee14416
https://suno.com/song/c3e633af-4893-4cd8-802a-2f5f42201a13

I'd be interested in seeing example songs where musical or emoji notation has had an effect.

[naive quail]

I didn't test them to that extent, but I found the same thing; nothing repeatable.

[stoic sleet]

Can you give me some tips on how to switch between male and female vocals in different parts of a song? Thanks in advance!

[daring lava]

Using things similar to “dialogic duet”/ “antiphony” / “Interlocutory Singing” in the style of music can help. In the lyrical field having it set up like such proves useful…

[male voice]
…..

[female voice]
…..

Please keep in mind that this could take some rerolls to accomplish. Still room to grow for sure!

[charred merlin]

Preferably, a simple A/B test or something that doesn't require listening to a full 4-minute techno breakdown with multiple verses, trying to isolate one instance where the symbol might have been interpreted. It's great if someone is getting consistent results using their methods, but it's especially important in a thread for technical discussions to be able to back up your claims if you're going to state your advice as objective fact.

Tags are suggestions, and not the same as lyrical scripting. It's like feeding code to the front-end of ChatGPT and expecting it to follow it precisely-- it might try to understand what you're saying, but you haven't "brainwashed" it to obey your commands unless you have access to the backend of how those generations are produced (like Custom Instructions or Memory, in the case of ChatGPT).

Here's a couple experiments, all with the style prompt "experimental";

• https://suno.com/song/add2b886-e7ea-42d2-b660-c5791fb16802 lyrics = "🐰🐇🎺♀️🎤 Testing 𝇟" >> no trumpet, male vocals repeating "testing" for 4 minutes
• https://suno.com/song/23dcdd3c-35f9-45e5-ae1a-b8272995fc8f lyrics = "🐰🐇🎺♀️🎤 Testing 𝇟" >> no trumpet, female vocals, 1:27
• https://suno.com/song/ba0cc12f-0912-4314-ab38-e0194951e630 lyrics = "🎺🎺🎺🎺🎺🐰 🐇🎺♀️🎤Testing 𝇟" >> no trumpet, male vocals, symbol pronounced as "goteh", repeating "testing" for 4 minutes
• https://suno.com/song/dbbdef69-2914-4fe5-bcd6-9dc569d62305 lyrics = "🎺🎺🎺🎺🎺🐰 🐇🎺♀️🎤Testing 𝇟" >> no trumpet, male vocals, fully hallucinated verses and chorus interpersed with "testing", 3:57

[raven turtle]

I really like the generally upfront, relatively good fidelity being generated by Suno, but I've noticed that the output feels over-compressed and borderline clipped to the point of distortion. I guess this may be either to get the volume up a bit, or the model being trained on modern, brickwall mastered content, but has anyone noticed this too?

[unique stone]

Custom/instrumental with only one of these in the style and no genre gave me generic piano: 🎸 🪈 🎺

I'll try some more things (including seeing whether the behaviour is the same if I just type out the instrument.) It does seem like putting ♀🎙in the style tag may work similarly to "female vocals" - if so that'll be a good way to save prompt space there.

[naive quail]

the art renderer is able to interpret emoji's in the lyrics, at least

[charred merlin]

This is a neat github experiment, but it's not the official Suno API.

If you want to get involved in that repo, you can submit feedback through their Issues tab, or make your own pull request if you want to improve upon it.

[unique stone]

There is currently no official Suno API at the moment.

The APIs you have linked to are unofficial developments/hacks that use the website in a way that is not intended, which likely violates the Suno Terms of Service so should not be discussed here.

This message, and the two similar ones you have sent in the Portugal/Brazil channels have been removed.

[lilac garden]

When i extend a song sometimes the lyrics are wrong or duplicated , is there a way to update or correct the lyrics

[cobalt zodiac]

Is there a tag i can use to make my songs not sound distorted? I dont want to use external programs to edit

[daring lava]

If you’d like to share your prompt, I can try to help refine it

[cobalt zodiac]

https://suno.com/song/8a510389-9751-4e85-a693-a0dc4ffb40d8
Basicly all in this playlist, Starts off great but for some reason the bass / Volume just gets increased and then stuff just sounds distorted and bad

[daring lava]

Could be a number of things, where you’re extending from or how you have the new text inputted. If you’d like to share what you’re working on it’d be easing to work out

[cobalt zodiac]

This happens especially if making House, Drum and bass etc

[charred merlin]

Out of curiosity, how many extension points does the song consist of? Roughly

[cobalt zodiac]

Happens regardless if i extend a song or not.

[cobalt zodiac]

2

or well 1 extension

First generation + 1 extended

[charred merlin]

Do you feel like the change in quality happens primarily after the extension, or does it noticably change even before getting to that point?

[cobalt zodiac]

Even before getting to the extension point, In the above playlist there is only 2 songs that have extension

I have realy good ears and bear with me here,

The bass is thumping hard and sounds nice, But it feels like there is something else added by the AI that in general just turns everything to 200%

(If that makes sense)

[charred merlin]

Kinda like smelling every candle at Bodyworks and going noseblind, I try to not overanalyze every track that people share & hope people are able to explain in detail what they've noticed, easier for everyone to compare observations rather than needing to build a "mental playlist" to compare everything and try to make sense of it.

Listening to your track, do you feel it gets louder when "the solo" kicks in @ ~1:20?

[cobalt zodiac]

Yeah i agree, Im a EDM and Rock/Metal head. In both genres i discovered this "distortion", I had to generate and generate to get the best results, At one point with the song https://suno.com/song/8a510389-9751-4e85-a693-a0dc4ffb40d8 i just had to give up. I think i went through over 100 tags and 1000 credits

I did download the songs as i thought Suno's music player was bad but it sounds the same. and call me on edge but it sounds like the songs degrade

[daring lava]

Very interesting. A couple things to keep in mind, still very early days here and lots of room to grow. If you’re noticing consistent results that you have concerns with I would most certainly post it in the #🪲┃bugs-web channel so the team will see it (don’t ping). Consider using something to better adjust the equalization of the sound in post via third party. (This is for non-extended versions)

For extended versions, be mindful of where you continue from. Avoid areas with high level of distortion when possible.

[naive quail]

I think that increasing distortion has been reported pretty frequently

[cobalt zodiac]

Yeah i learned that pretty fast, Have to be picky from where you "extend". I will make a post in #bugs

[cobalt zodiac]

From a quick google search there is many reddit threads regarding this

[naive quail]

I notice it much more frequently when extending, but definitely hear it in untouched songs, too

[cobalt zodiac]

One of my summer vibe takes:
https://suno.com/song/f95416fc-157b-4dfd-82ce-17aa64d520c4
The bass line at 0:50s seems to be cranked to 200%

[daring lava]

Please share what you make in #🐣┃suno-showcase not here in technical discussion unless of course you need help with it. Thanks

[cobalt zodiac]

Yes that is fine, The links above are directly linked to my previous post and are given as samples since its related to the discussion at hand.

[charred merlin]

Whoa-- can someone test something for me?

#📚┃suno-school message

The 'Extend' button shows up, but are you able to extend any of these songs?

(all set to private)

[daring lava]

I’ll give it a shot

[cobalt zodiac]

Yep, Press the 3 dots and i can extend it

[charred merlin]

If you click extend, does it actually let you?

#🚀┃updates message

NICE

[naive quail]

Yea, extend doesn't care whether it's public or private

[charred merlin]

OH MY GOD

[daring lava]

Yes I can.

[charred merlin]

I'm gonna take my excitement elsewhere because it might be a bug report instead 😛

[naive quail]

probably (hopefully?) falls into the bug category.

[charred merlin]

Thank you for burning credits for science. You're a real one ❤️

[wintry tiger]

does anyone have a song that uses the lyrics but it has to be instrumental only.

[faint crypt]

Is there an app coming into he future

[charred merlin]

Not sure what you mean-- are you asking if we've ever received lyrics in a song set to Instrumental Only?

[wintry tiger]

it's like that but I want to control when the instruments kick in.

[hollow hare]

Or do you mean people using instrumental tags only to make instrumental songs?

[hollow hare]

You will want to use instrumental tags.

I like using RayanAIR's Novo Lands as an example to show how flexible these tags can be

[wintry tiger]

thank you.

[broken lotus]

yes, Suno put out a call for ios and android devs... not sure the current status of either.

[faint crypt]

That's amazing

[undone kelp]

Hello! Anything about this distortion/overdrive? It ruins all my songs (maybe not just) today, Suno's masterpieces. v3.5 Focus on the third verse (chorus) in it, btw it's always a worsening gradient; the song begins with a relative good sound quality but collapses as we reach the end.

[vestal rapids]

what is the time stamp of the issue?

oh yeah I hear it

[vestal rapids]

nasty

[graceful condor]

diffusion or language modeling for suno? whats your take on what they use?

[vestal rapids]

Language Model, I used to think diffusion model but people on here convinced me otherwise

[graceful condor]

my same feeling, bark seems to be the groundwork for alot of it along with some internal customization. Diffusion models like stable audio open are great sonically but they just lack coherent song structures and creativity

[worldly mauve]

I would say diffusion

but it's not "diffusion or language model"

it's "diffusion or autoregression"

and I'm relatively sure it's diffusion

like the fixed clip length and context size speaks for diffusion

but it could also be a mix, like using autoregression to generate "musik tokens" which then are translated to music via diffusion

[solar rose]

it's amazing how often suno doesn't recognise the difference between male and female. I want a male voice for my current song, but no matter whether I enter male singer or male vocalist or both, it keeps generating a female voice that sounds completely overdriven.

[solar rose]

Unfortunately, this distortion or "overdriving" occurs quite often, especially with female voices in my opinion. Sometimes you have 3 good songs and then the next 10 are overdriven again. It's very annoying when it generates a song that you really like and then ruins it.

[undone kelp]

Ah, I mostly hire male singers. 🙂

[solar rose]

I do that too, but suno very often ignores that and delivers a female singer 😫

[unique stone]

It can be difficult to get Suno to listen to exactly what you want. If you haven't already, try repeating the same part of the prompt several times to basically nag it a bit 😄 so like "male vocal, male vocal, male vocal" and try putting that toward the start of the prompt.

[graceful condor]

when I say language model, im implicitly referring to using audio discrete tokens

and I dont think its diffusion, if you inspect the spectrograms you can clearly see they are coming from a neural decoder

they might use diffusion as a conditioning for some process but final output is codec

[graceful condor]

and to be fair, language models can be non autoregressive too 🙂

[worldly mauve]

yes, that's why the original question was misleading

of course it's also a language model, it contains song texts. Question is if its a Autoregressive model with discrete tokens or diffusion or a hybrid of both

[graceful condor]

im referring to the output obv, you can use lms just for embeddings

[worldly mauve]

I cannot imagine how you can generate a coherent song in a purely Autoregressive way. You have to process the tokens of the complete song afterwards, transforming them into music

and the most efficient way to do that is usually via diffusion

[undone kelp]

I once got a beautiful conversational duet. It could have been beautiful if not ruined by distorted overdrive. 😄

[worldly mauve]

just works better than gans or vqgans and similar alternatives

[worldly mauve]

the embeddings from language models are just coming from stacked transformers. In the end all models, be it image generators, music generators or text generators are using transformers nowadays. The question is what the model is doing with the embeddings to produce music

[nova hound]

Is itvpossible to bulk download all songs we made?

[charred merlin]

The only reason I lean toward thinking it’s Diffusion is because of the hallucinated vocals in Instrumental mode; the “lyrics” are really simple rhyming syllables that seem like they would be “diffused” into preferred lyrics if they were entered, or what should have been re-interpreted as an instrumental lead (but was erroneously diffused as vocals)

Or at least because image gen is the closest comparison I have to working with generative AI firsthand

[broken lotus]

No, but that would be so useful!

[vestal rapids]

so would a download button, so you don't have to dive into a menu to download, that's really poor UX

[wild shuttle]

does the "extend" feature only listen to the audio you tell it to extend from, or does it take into account its metadata (the prompt, and whatever else suno saves about the generation)?

[broken lotus]

the strongest influence is what it can hear at the end of the earlier clip.

because the audio from before is so strong, there are some rules of thumb when to get whole song and extend from that, or to create a really short clip that doesn't have much memory so the style is very easy to change...

https://www.suno.wiki/faq/making-music/extend-from-time/

[wild shuttle]

thank you! i don't want to change the style, on the contrary i'm more wondering how to keep the song consistent through all the extends

and i'm curious about the technicalities

because if it's just listening to the audio, then is it the same tech that would be used with uploaded audio?

[worldly mauve]

that's why I guess it's diffusion 😛

diffusion naturally allows for audio2audio

[nimble pulsar]

Some of the songs I generate go on too long, especially if they have been extended. Is there any way to trim them to a certain spot in the song?

[celest compass]

I've been having this issue all day.
Lost a lot of credits today with nothing ending. I've been using Suno for a couple of months, so I'm used to trying different previous song end points to work an end point (with various combinations of End, Fade, Fade to End etc)
But today, even on a perfect quiet point in a track, I try to extend to end at this point and it then goes and creates another 2 minutes.
It's felt the first time I'm been frustraited today with Suno. I never mind losing the occasional credit getting an end point, but I'm on three different tracks and must be way over 40 different attempts to end today.

Oh... Let me go try changing to v3 to see if that helps.

[teal raft]

I have no issues doing an ending with v3

I know 'works for me' and all that

But it's so reliable for me that I take it for granted that it will work, you know?

[fluid robin]

[Long tuba Solo]

[instrumental interlude]

[Long Trumpets Solo]

[instrumental interlude]

[long Drum and snare solo]

[hollow hare]

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

[celest compass]

I've just used the same 'ending' that wouldn't close out on the 3.5 songs.
And I've just closed 2 down within the either the first or second attempt. 👍
So, maybe it's the workaround for the 'ever extending v 3.5', change the extent to 3.

[teal raft]

That's the current accepted way, yeah

Switch to v3 and scream 'END' at Suno a whole bunch

[celest compass]

Which is what I was doing (at least 3.0 listens) 🤣
Thanks for clarifying this . 👍

[graceful condor]

musicgen does this in a ar way

[graceful condor]

right, but you would not call stable diffusion a language model, its really just a 2d CNN based unet conditioned on embeddings. I am referring to what is used during the generation process. Diffusion models suffer from poor sampling time, using discrete codes is much faster. Unless you use something like stable audio did but I doubt suno trained their own VAE in the raw domain for diffusion

[graceful condor]

this doesnt have to mean diffusion, just some form of conditioning

[graceful condor]

any sort of conditioning allows for this. Check VALLE

[worldly mauve]

stable diffusion is not really a 2D cnn unet 😬

yes, it has cnn, like 1% of it's parameters are a cnn. 90% are transformer layers

[graceful condor]

how is it not? the latent embeddings are fed to a 2dcnn where the diffusion is performed. Most of the parameters are actually in attention layers, but they arent really transformers

[worldly mauve]

yes, but you have to train conditions explicitly. Diffusion allows to do audio2audio without training for it

[graceful condor]

attention in 2D space can still be attened to, its not a transformers. Transformers work on steps with D dimensionality, you can use a VIT but SD is not a VIT

[worldly mauve]

of course they are transformers... what's the difference?

[graceful condor]

using attention does not mean it is a transformer

[valid flax]

Yo why i cant suddenly change position of the song playing?

[broken lotus]

It needs to fully generate. it spins while streaming in lesser quality.

[worldly mauve]

I would say the main differences between a ViT and SD is that a ViT has positional embeddings while sd doesn't have them

[valid flax]

no the UI is bugged af i am gonna record the behavior 1 sec

[graceful condor]

so wrong, a VIT processes the image as a sequence in patches, a CNN process the image with a sliding kernel, attention is only used to get over the fact of the receptive field of CNNs.

[worldly mauve]

not a big difference

patching and cnn works very similar

actually you can represent patching as a special case of cnn

[valid flax]

What the hell? XD

Firefox btw

[graceful condor]

computationally, convolutions are much faster

[worldly mauve]

no, why would they be faster

[mystic fable]

Does anyone know the best way to make suno more consistent with sound effect tags?

[worldly mauve]

it's really the same

[valid flax]

I cant forward songs

its just me?

[worldly mauve]

in both cases you split the image into patches or into smaller resolution and apply transformers on the latent pixels

[graceful condor]

its not, cus VIT is using attention over the entire sequence. Using a CNN doesnt mean you use attention

you can pick and choose where the attention goes

[broken lotus]

I'm not seeing that, but looks definitely like a https://discord.com/channels/1069381916492562582/1082775596767117312

[worldly mauve]

I didn't say cnn is the same as ViT

[graceful condor]

i said CNN with attention is faster

because you can control where the attention goes (middle layers, etc)

[worldly mauve]

in the unet you also do attention on thr complete image

[graceful condor]

please read the paper i think im done

[worldly mauve]

I know the paper

and I know the code

[graceful condor]

I dont think you do but its all good

[worldly mauve]

it's really not that different. SD is a transformer based architecture. Yes, it's using cnns in between.

main difference to a vit or a dit is that it's not using positional embeddings. But attention is applied on all the latent pixels

as in vit where you apply it on all patches

[graceful condor]

its not though, because you can process N channels in parrallel with a CNN. in a transformer you have to extend the sequence by N

[worldly mauve]

I mean you don't extend a sequence in an image

because it's a diffusion architecture and not autoregression

[graceful condor]

? you have no idea what your even saying. a CNN can process all channels of an image at once with a kernel, a transformer has to flatten the image (Extend the sequence)

dont confuse new people who are coming here, please know your stuff

new people who will read this will be confused by your statements lol

[worldly mauve]

actually I also don't understand what you mean. A transformer is also applied on all tokens in parallel

[graceful condor]

tokens = sequence

only way to create a sequence form a 3dimensional image is to flatten it

[worldly mauve]

I mean a transformer doesn't operate on sequences but on sets

and an image is a set of pixels

[graceful condor]

lol

[worldly mauve]

flatten is a zero-cost operation

[graceful condor]

just think about it. We can break it down to a 1d conv if youd like

with 1D, it processes a (Batch, channels, steps) and a transformer can process the same sequence using (Batch, steps, channels) right? so now lets bump up the dimensionality to 2D. you got (Batch, channels, height, width). for each position in a CNN kernel it can look at all N channels in parallel. Now for a transformer to process this, you would process it this way (Batch, height * width, channels). Where in the 1D scenario, we dont incur any additional steps to process the sequence. In the 2D scenario we do

[valid flax]

Looks like suno is completely broken on Firefox atm

Reported with all relevant stuff

[worldly mauve]

"Now for a transformer to process this, you would process it this way (Batch, height * width, channels)"

Yes. That's what's done. So what?

[valid flax]

By the looking at the console logs ... I assume some analytics stuff wont get loaded due to invalid CORS. Which sucs because some functions from their analytics gets called upon track position change which fails

[graceful condor]

you process it? what happens when you run transformer on a sequence thats triple or double the size? it takes longer.. there is less parallelization going on

[worldly mauve]

yes, transformers are slower than cnns - I never doubt that

and yes, the size of the sequence grows quadratically in images

that's why you have to use patching and/or downscaling layers, otherwise applying transformers on full resolution would be too costly

but it doesn't make a huge difference if you use patching to reduce the size of your sequence or if you use a downscaling cnn for that.

[graceful condor]

well my point was mainly about CNNs and complexity and why they arent transformers... even in the 1D conv the complexity is o(kernel_size * steps * channels^2), like you said self attention runtime is O(n^2 * d), so runtime grows quadratically with sequence time. You made it seem like additional steps werent a big deal, but trust me if stable diffusion was a transformer it would not run on a 3090

[worldly mauve]

I know that CNNs are not transformers. My point was that SD IS A transformer framework

just because it contains some CNN layers doesn't make it a CNN

as said, 90% of its parameters are transformer weights

1% are cnn weights

and it performs normal global attention across the complete image

as a ViT

[graceful condor]

what are "transformer" weights lol

there are not linear layers in DF

SD*

[worldly mauve]

the V,Q and K matrices as well as the MLP of the transformer

[graceful condor]

thats self attention

does not mean transformer

[worldly mauve]

no

there is a linear layer

[graceful condor]

no, I mean in a transformer it uses linear layers BEFORE and AFTER self attention. using linear layers to compute self attention is still just self attention. read the transfomer paper

[worldly mauve]

yes, I know

I mean you are right in the sense that transformers usually have two linear layers that go from low dim -> high dim -> low dim

in SD it only has one linear layer

but it has a linear layer BEFORE and AFTER the transformer

[graceful condor]

thats called a projection layer, usually just linear doesnt even use activation. the ones in transformers use GELU

[worldly mauve]

yes, I know. This is indeed a difference to transformers

in SD you have that, too, but it comes after a block of transformer layers

I mean it's not an insignificant difference - cause the MLP block is often the most important part of the transformer. Still I find it a bit weird saying it would be no transformer but "just attention". Anyways, at least I know get what you are referring to

[graceful condor]

I say this because people initially used the concept of attention on seq 2 seq probelms in recurrent nets, as well as 1dcnns. Transformers specifically refer to the architecture that utilized attention along with the other components (layer norm, etc) to be as expressive as they are. using attention alone doesnt guarantee the same capabilities. Theres a reason they are transformers and not just attention

[worldly mauve]

yes, but this is not the case here. You have everything that a transformer should have

I also checked the code and was wrong: every transformer layer is followed by a resnet layer

so although the exact architecture is different from transformer (no MLP with GELU but a ResNet layer with SILU) it's conceptionally very similar

[graceful condor]

its a 2d unet with a attention, kapeesh? lol

[worldly mauve]

I see we don't agree here lol

[graceful condor]

it seems like your just not reading the papers or something lol

theres no doubt that the same components exist but they are not the same things at all

[worldly mauve]

dude, your "you are stupid and I'm brilliant" attitude it's really annoying. We can disagree on certain concepts, but I have a phd in computer science and YES I read the paper

[graceful condor]

the high frequency features actually come from the cnn encoder skip connections, did you know? attention just helps aggregate those features. Im sorry but no where online says that stable diffusion is a transformer lol your making this up. we talking about 1.5, right?

your just making up information though

read the paper, post the arch,

[worldly mauve]

boah, I read the paper. How often should I repeat that

[graceful condor]

this is alarming

i mean it says its a unet, im not just making up information

[worldly mauve]

I think its nitpicking when you say something is not a transformer because its slightly different from the original transformer paper. It is a attention mechanism followed by a mlp with nonlinearity as in the original transformer paper, even if the detailed architecture is a bit different

[worldly mauve]

yes, it's a unet. And yes, unets were originally pure cnn architectures. Seems to confuse people. maybe that's why they nowadays call them "hourglass transformers" instead

[graceful condor]

stop confusing other people though, this stuff is right there on paper

[worldly mauve]

I don't confuse other people lol

• SD is not a CNN architecture is a correct term. It is not a pure CNN
• SD is using transformers is also correct, even if you call them "just" attention

[graceful condor]

lol ok lets just revisit this

dude it says its a unet using attention, if someone wants to build their own model to replicate this work lets say on hugging face they would use a 2dunet arch...not a transformer architecture..your not sane in what your saying so just refer to the paper we dont have to make up information

so lets steer them in the right direction at least

theyll be out here trying to use t5 for image generation

smh

[worldly mauve]

t5 is a good example.... Like the original transformer used encoder-decoder architectures that are rarely used nowadays

still nobody would say llama is not a transformer because its decoder-only

anyways

yes, if people are not understanding concepts and just looking for names

I still don't understand how I would confuse them saying its a unet with transformers, but anyways

[weary wharf]

🍿

[graceful condor]

lol honestly i did not know if you were/are trolling. just did not want to confuse people who are new to the space

[worldly mauve]

I'm not trolling. I just think its nippicking to make a difference between "unet with attention" and "unet with transformers". Even in their code it's called transformer everywhere. But anyways

[graceful condor]

ok sure if you want to land on "transformer encoder" im with it, there is a reason a DIT is a DIT, but sure

[worldly mauve]

I won't say a DiT is the same as a Unet with attention - I just think it's not as different as it sometimes is stated. DiT using patches and unets use cnns with downscaling is not a huge difference at all. I agree, though, that the transformer math itself is different between both and maybe this makes a difference.

[weary wharf]

Anyone trying neo-classical and/or dark-wave for style in v3?
It almost comes out as "operatic" and not that "cinematic".

And styles that you found succesfull?

[worldly mauve]

and for sure the DiT architecture is more elegant

(I rather suspect that the use of positional embeddings makes the difference. I always found it weird that they left them out from the unet attention)

(and other implementation details as the adaptive layer norm and so on)

[graceful condor]

in my opinion its scale... training transformers they overfit easily. you usually need lots of data. CNN archs can get there faster initially, but then its a slow climb

so directly related to self attention

CNNs do not have the same size receptive field unless you use dilation, and has not worked as well in practice

[worldly mauve]

the thing is that the unet with attention architecture has cnns but also attentions + mlps with residual connections. So even if the cnn layers might have a small receptive field, the attention layers have a global view on the data. The problem is: without positional encoding the pixels in the image cannot attent to each other in any useful way cause they don't know how they are located to each other.

when I first read through the SD source code I thought this is a bug and they just forgot the positional encoding. But it seems that the CNN can somewhat take over

so the CNN has to do all the compositional stuff and as you said it doesn't scale well on more training data

if you look at the hourglass transformer architecture it looks very similar to the unet, just that they removed the cnns entirely. I guess you don't need them anymore if you use positional embeddings

anyways. I think I agree with your opinion and we might just argued over namings which, I think, is a stupid reason to fight lol

just to my defense: people very often say "SD is a unet cnn architecture" and for me that's as wrong as for you hearing a unet is a transformer architecture. The original unet paper was a pure cnn without attention at all, but that has nothing to do with the architecture SD is using and people often confuse that.

[graceful condor]

yep, CNNS embed time information implicitly in the sequence. like i said I thought it was a troll but yea nitpick namings I see your point. Just trying to steer other people who would want to replicate in the open source and learn the foundations. Alot of time when people hear transformer they automatically think of GPT, when we havent even discussed transformer decoders so it can get confusing for people at first glance

[graceful condor]

I get that, but having worked alot with audio and imagery unets have been vital to compression and generation so they are the real foundation on the architecture whether it has attention or not. in my experience attention is more like bells and whistles, unets are already very capable and like you said using periodic functions like siren can implicitly embed a positional encoding to help overcome small receptive fields, not using attention at all.

attention is great for using information (conditioning)) but no reason it could not be done be concatting along a channel in a cnn. many ways to condition

[worldly mauve]

that might be the case, but in SD it's really the attention layer that do most of the work. When people train the model on new data (new styles, new concepts, new objects and so on) they usually only train the attention weights

[graceful condor]

I dont doubt its vital in SD, just not the only way. plenty of pure conv GANs have proven them selves, they are just hard to optimize

[worldly mauve]

I also won't disagree with that. But nowadays people are often like "it uses transformers so it has to be better" 🤷‍♂️
(and, scarily, quite often they are even right with that 🙈 )

[graceful condor]

lol facts...definitely comes down to scale sadly...more money better models usually

[hollow furnace]

I'm hesitant to thumbs-down songs while generating even though it's the best way to clear out the library because i don't know if this is fed back into training data. Does anyone know?

[broken lotus]

yes, but not directly.

Suno uses the data (especially thumbs down) to get feedback from us, info they use to train the next model.

[unique stone]

Thumbs-down does not remove the song from your library (it just doesn't show on the list of songs when using "create"). I have a bunch of thumbs-down'ed songs in my library still because I'm messy and don't take out the trash 😅

[sleek aspen]

how would you make a piece of classical music 10 minutes long?

[weary wharf]

I'd try a few styles in first with 3. Maybe something like this:
classical, folk, soundtrack, world, middle-eastern-instruments, emotional, journey, transformation
or
film-score, movie-soundtrack, symphonic, orchestral, symphonic, orchestral, intense, desperate, betrayal

If it works, I would try 3.5 to see if it's capable of sticking to the genre and style. If it's, you can use 3.5 to extend a few times, to go to 10 minutes.
Else, stick with 3, and do 2 + 1 + 1 + ... until you have all you need.

But I'm sure this will burn too much credits.

Here is another one
film-score, movie-soundtrack, western-classical, symphonic, orchestral, classical, powerful, dramatic, expressive, spirituals, vibrant, adventurous, dynamic, emotional, epic

This is too long for style input I guess. But you can cut.

[broken lotus]

keep it evolving
https://suno.com/song/f96cb9bd-133f-4f61-ab22-e21dc736eea2

[hollow furnace]

hmmmmmmmm, seems like then, there should be an easy to hit "x" button for delete (right now it's really fiddly and slow) or they might get some pretty misleading data from people just trying to keep the Create window under control? (even if it doesn' t remove from the library, hmm)

[patent fox]

So tags in lyrics like [Verse] and [Chorus] work. Are there any others I should be aware of, or that are useful?

Oh it's literally int he pinned comments lmao

I guess my question would be if you can double up the tags like [Verse][Belting]

[broken lotus]

[belting verse]
or just
[belting]

[modern belfry]

question. Is there a way to add lyrics during a solo ? It'd be kick ass to be able to do that in a song I'm writing

[opaque jetty]

Is there a way to do differing vocals and then combining them? For example if I want to have a song where say a female vocal does Verse 1, then a male during Verse 2 and combining them in a harmony/duet for the Chorus.

I've tried:
[Verse 1 Female Vocals]
[Verse 2 Male Vocals]
[Chorus Male and Female Vocals]

That and many other variants will work sometimes, then during the course of the song the AI will just stop caring and do whatever. Male will be female, the harmony will go away, etc.

[patent fox]

I've had some luck with adding "backup vocals" and "female vocals" to the music style, and adding the male lyrics in parenthesis. But a lot of the time I came across the same problem as oyu

[daring lava]

Have you tried extending it from the part right before the chorus part and adding duet to the style of music field? Of course you’d need to adjust the lyric field to begin at the chorus, but that could help. (Haven’t tried this so experiment at your own risk)

[soft lynx]

Not sure if we can link to other songs, but check this one out which does what you are talking about. https://suno.com/song/338ceeaa-dbc7-4f64-9689-89e8bdea92f3

[sour tapir]

Hey there 👋🏻 Is there any meta tags for suno to create a wicked intro (for eg. W w welcome to Suno AI) ?

[sharp kraken]

i do duets like that and the way i do it goes like this #1 include both vocals instyle i.e. "Soft female vocal, Male Punk Vocal, Duet" - then use those exact tags on the verses i.e. "sift female verse 1" etc... then include in the description when you want them both to sing together i.e. "[Chorus 1: Male and Female Duet. ] "

[opaque jetty]

@patent fox Thanks for the insight, glad to know I'm not the only one 😉

@build.a.better.world Hmm... interesting idea. No I haven't tried that yet but it sounds crazy enough to give it a shot. Gracias!

@soft lynx Woah. That is an accomplishment, wow. Yeah it looks like the AI did the changes on its own, I don't really see any prompts or anything to force it. Maybe if I make my song a 10 minute-r it'll be more nice with my suggestions 😆

[soft lynx]

previously it did have the changes in it.... as in [country voice] [rap voice] or something like that. maybe reach out to @seticamel

[graceful condor]

pro tip, follow lyrics annotations for variable voices (genius)
ex - [Verse:Mr Krabs]

[tribal flare]

Elaborate more please, could you have example?

[echo sundial]

Hello to every one, just a question :Is it possible to change small parts of the text (to correct) while maintaining the same arrangement? or do I necessarily have to make a new song?

[prisma wraith]

You could continue extend* the song from a bit before the mistake, it will cause some changes to the song, but it won't change as much as if you generate a new piece from scratch

[iron sparrow]

Heya,
Does anyone have experience about creating a new person using a speech to text, then using that person as a text to speech?

Details: I would like to use the counter strike voices to comment actions happening in the game I am building (so I would download the counter strike voices and feed that to create this new agent, then make the agent read the actions)

Or using a text to speech but with a prompt to describe the user? So that I could write in the prompt "the agent has some robotic voice" etc

[polar abyss]

its a real struggle... seems like the order you present male/female in the prompt makes a different... so you sometimes have to re-organize while extending. When you want the duet you need to tag with [Duet] and just spam rehashes... Sometimes just using call response with parenthesis will also give you the alternating, but at a faster rate...

Here's one I did... you can see my spammed tags in the lyrics... but i changed the order of things in prompt A LOT during extends

https://suno.com/song/f17c0ca1-e3dd-4906-8ead-455e435a71a3

[steel cradle]

I’ve done duets just by specifying different voices in notations

[polar abyss]

The complicated part is controlling who is singing as you continue to extend

It will ignore body tags at some point so there's some weird finessing you end up having to tinker through

[steel cradle]

Indeed, I don’t know too much on it as I started relatively recently, but I’d say I’m not too bad at prompting it, especially in custom style and lyrics mode, I just thought I’d point out that it is possible, even if not practical

[opaque nebula]

Hello everyone, I want to create music, where first part sings woman, another part - man, third part - together. Can I make it using suno.AI? I have tried to write [male voice] in style of music and [male vocal]/[female vocal] in text of music, but it doesn't work!

Like here: https://suno.com/song/d39a6bbb-eca2-471e-b56c-158fb59e3469

[broken lotus]

I see you are trying a few things at the same time..., maybe simplify...?
the bot sometimes does best with 1 short metatag:

[Judy]
Well, Hi my name is Judy!

[Steve]
And I'm her boyfriend Steve

[Judy and Steve duet]
There's a chance we are singing together now

[opaque nebula]

Oh, Ok, thanks, I will try!!

[floral sonnet]

hi, i don't understand something... does MrTree upload sound effects into Suno? .. how do you generate specific sounds like 'gunshot' or 'whip' ...?... also is there a list of commands/ prompts somewhere ? ... i suppose I've got to do a learning cycle on youtube or 'summit...can anyone help me?

like this --> https://suno.com/song/a51a6268-3194-4e8e-86bc-7a7ea7955f51

[naive quail]

The ability to upload our own audio sample for Suno to then "Expand" upon is coming with v4. Some people have early access to the new feature for testing/debugging purposes.

[sharp kraken]

i have no idea how mr.tree makes his tracks. i have re-used his prompts but it doesnt sound the same

[floral sonnet]

ok ... can i get early access please 🙂

[alpine furnace]

with extend and a number of credits

that related to the 10min questions sorry slow update of the caht

I have a technical question - how do i modify the spoken verse pitch to be very deep almost rumbling

[Very Deep Male Bass voice Verse 4]
this didnt cut it

[floral sonnet]

@sharp kraken you are doing some crazy cool stuff too !!... that seems outside the realm of normal prompts.... i like your grim. stuff...

[sharp kraken]

tank you. i use alot of unicode symbols intended for music composition

[charred merlin]

Demonstrate your willingness to help Suno grow, be an active part of helping out however they ask (properly submit bug reports on Discord, use the bug reports & thumbs down on bad audio quality generations, etc), and they might take notice!

A lot of people unfortunately view early-access as a way of getting the cool new stuff immediately, but don't have the mindset of helping. (Like when you sign up for beta iOS, don't call Apple for your issues, you've agreed to submit bug reports through the beta feedback app! 😛 )

[floral sonnet]

@charred merlin I beta test for Melodyne Celemony, we got a Grammy for best technical software!...and I beta test for the following companies: Waves, Fabfilter, Galbanum, SFX Machine, Olympia, Audio Ease and Peterson Tuners... who is it that takes notice ?...

[charred merlin]

You could email them if you feel like you have something to contribute, [e-mail], but I was talking about getting noticed by being actively involved with Suno, not other projects ❤️

[broken lotus]

The members who got early access got it through a Twitter promotion...
(not sure the employee that made that happen is here on Discord)

[charred merlin]

Not the only way I'd imagine, though

[floral sonnet]

i'm playing around with one of your prompts but i notice that <6/8> doesn't really work... like i ask a specific time sig and it generates in 4/4... its amazing that you get it to go so dark ... why is it that whenever I generate something it sounds so 'happy' !!! <-- not haunted at all!! .. i want it doom-tastic!!

[sharp kraken]

6./8 doesnt work because its 3/4

[floral sonnet]

thanks.. I'll try support..

[sharp kraken]

i meant to do 6/7

you gotta have the right adjective i battled it forever to quit spitting that "pop" sound

[floral sonnet]

yes.. i mean the time signatures don't always work... it doesnt 'force it' to do it in 7/8 ot 3/4 etc...

[sharp kraken]

I always assume it was trained off of spotify - because it had no idea what "joshuascsca" was - so it works better if you use spotify tags

yeah its never forced and sometimes you have to wait for it especially if it's command cue gets all backed up

thats why I try to space commands out

[floral sonnet]

good name BTW 😉 - im going through your prompts at the mo.. intense!.. seem to work! i like this -> https://suno.com/song/b27de133-7b5d-45fc-b714-1c377d91b43f

[sharp kraken]

check this one out https://suno.com/song/ee34c776-1811-46fa-8704-00b8fbc98867

more like a tutorial\

than a track

[charred merlin]

The different generations for Language of Music all seem wildly different from one another, with only the lyrics seeming to be the only common attribute they share, though.

Using the same symbols and beat notation, I would expect them to sound similar (if Suno respects them the same as the tags listed in the suno wiki)?

[floral sonnet]

@sharp kraken why does it sound so happy if i re-use the prompt >.?≥. is it because most of my generations up to now have been emo, indie ,melodic , pop songs? ,,, like my algorithm is infected!! 😦

[sharp kraken]

You can change the scale thats probably the best way

most of my music is evil scale or octatonic scale

"evil scale" is not a musical term and im not sure where it came from - but it works nicely

im still messing with note length prompts like these "𝅘𝅥𝅲-𝅘𝅥𝅮" they control the note length and they work along with a sort of medieval version of rhythmic annotation. They affect note length and the .-'s affect pause length

and you can of course create a fixed melody with actual notes using modern notation like C#6

[floral sonnet]

Ive tried 'microtonal' scale and it seems to work on other peices... evil??,... not yet.. but will now 🙂

[sharp kraken]

yeah and it also seems to fully understand gregorian notation - which is like obsolete for centuries - but extremely easy to compose

aand possibly more flexible because it doesn't require specific note

[floral sonnet]

WOW!!! you can put note in like C4 etc ?... interesting .. (currently listening to your Spaced piece)

[sharp kraken]

yeah look

[floral sonnet]

*playing .. so if there like a list of command or can you just almost use anything ? experiment?...

[sharp kraken]

https://suno.com/song/7d17f718-b5cb-481b-91bd-8fa210a03b88

so this one uses both rhythmic and tonal notation

so nomatter what you do the melody and rhythym stay the same

its like a testament to my love for Hendrix Harmonics, the whole thing is built around the Hendrix chord

[charred merlin]

https://www.suno.wiki/faq/metatags/
Here's the community-curated wiki page of tags that Suno recognizes. Other uses claim to have specific control over aspects that a lot of us aren't able to replicate, and have debunked with basic testing earlier in the chats, but the absolute key is experimentation.

It's an evolving product, and your direct feedback (thumbs down & reporting bad audio generations) helps improve the model and level of control for everyone! ❤️

An easy thing to experiment with is hyphenating your prompts; try "evil synth" and "evil-synth" without a lot of other prompts to see what kind of differences you can get just from that!

[sharp kraken]

R.I.P. jimmi - best guitarist ever

[floral sonnet]

@sharp kraken -- i think you;ll like this iteration 🙂 ---> https://suno.com/song/49164a9b-e5a0-4c22-98cb-4556fc215caa

[sharp kraken]

If you prompt this it will always be the same

[Triads: C3 E3 G#3 ._𝅘𝅥𝅰𝅘𝅥𝅯𝅘𝅥𝅰𝅘𝅥𝅰𝅘𝅥𝅯.| D3 F3 A3 𝅘𝅥𝅰𝅘𝅥𝅲𝅘𝅥𝅲𝅘𝅥𝅲𝅘𝅥𝅲𝅘𝅥𝅲𝅬 | D#3 F#3 A3 -.𝅘𝅥𝅰𝅘𝅥𝅯𝅘𝅥𝅲𝅘𝅥𝅰𝅘𝅥𝅯.. | F3 G#3 B3 𝅘𝅥𝅲.𝅘𝅥𝅰𝅘𝅥𝅯- | C3 E3 G#3 𝅘𝅥𝅰𝅘𝅥𝅰.𝅘𝅥𝅰𝅘𝅥𝅯. | D3 F3 A3 .𝅘𝅥𝅯𝅘𝅥𝅰𝅘𝅥𝅲𝅘𝅥𝅰- | D#3 F#3 A3 -𝅘𝅥𝅲𝅘𝅥𝅲.𝅘𝅥𝅯. | F3 G#3 B3 .𝅘𝅥𝅲𝅘𝅥𝅲-𝅘𝅥𝅰𝅘𝅥𝅯 | F#3 A3 C4 𝅘𝅥𝅯.𝅘𝅥𝅰𝅘𝅥𝅰.| G#3 B3 D4 𝅘𝅥𝅰.𝅘𝅥𝅲𝅘𝅥𝅲- . ]

[fleet gust]

hey, I want to add these tags to a song
[Intro]
[Begin with Ambient-Pad, setting a dreamy, ethereal tone]

but it keeps singing them as lyrics. Anyone knows how to do it?

[floral sonnet]

@charred merlin thank a lot!! thats what i need 🙂

[sharp kraken]

Oooh that does sound cool - ill follow you to see what you come up with

[charred merlin]

It's always fun to hear people's wild songs in here, but if you come across a really neat discovery that you think we could benefit from, it's really helpful to share a simple track that demonstrates it (only the tags and style prompts required, with a small-to-medium lyrical sample size so we can see how the normal repetition bugs play into it as well)

Have fun experimenting! I'm sure your experience edge testing all of the other audio software will come in handy 🙂 🍻

[floral sonnet]

@sharp kraken 'spoin' is meant to be 'spin'.. 😮

[sharp kraken]

[Introduction: Begin with Ambient-Pad, setting a dreamy, ethereal tone. ]

try that gorilla

and make sure ever phrase you prompt is punctuated

[floral sonnet]

its pretty good with vocalisation.... i programmed with something called vocalizer years ago...and then got a hacked 'speak and spell' that my mate peter kember of spaceman3 made (he went on to produce MGMT).,. anyhow.. its amazing the vocalisation power of Suno....

[sharp kraken]

suno is kinda obsessed with both punctuation and capitalization

[floral sonnet]

hmm.. ok

[sharp kraken]

My crazy vocal focused stuff is over heree
https://suno.com/playlist/123e2214-2787-4af6-95b7-51c75c3744a4

[floral sonnet]

@charred merlin @sharp kraken you can see how rubbish i am at typing... there is no way to edit something afterwards if you make a spelling mistake in the lyrics is there ?

[sharp kraken]

https://suno.com/song/5f9755b2-80aa-4eac-8f69-de76711baaaa

This one is pretty fantastic Eng/Jap/Tibettan its based on the random glitch in GPT 4.0 where it tells you something cryptic in tibettan

[unique stone]

Which aspect of this should always be the same?
I assume this goes in the "lyrics" box?

[sharp kraken]

the rhythm and the pitch. The first part with letters is key notation, the seccond part is rhythmic notation "_𝅘𝅥𝅰𝅘𝅥𝅯𝅘𝅥𝅰𝅘𝅥𝅰𝅘𝅥𝅯"

i made 4 different versions, all the same. it can sound different if the tempo is different but its just faster or slower versions of the same triads

and some of the rhythms are broken they dont add up toa whole bar so i didnt quite get it perfect yet

but despite what my (one and only) hater says, rhythmic notation is understood by suno - it's just weird because its hard to understand scaling

suno also understands gregorian notation

and you can just use pictures of instruments like this : 🥁 🎻♀🎙

♀🎙 < 4 character / Female Vocal <- 11 characters

[fleet gust]

thanks, seems random to me. In some versions it adds them as lyrics, in others not

[unique stone]

I tried the instrument pictures and the male/female vocal pictures yesterday without any success.

[sharp kraken]

do you use windows or mac?

they have worked 100% flawless for me

[floral sonnet]

@sharp kraken some of the glyphs/picts dont show up on my computer... i just get little boxes with hex codes in them... 🙂

[unique stone]

I get different results for this latest "triads" example prompt each time:
https://suno.com/song/0987c222-feed-42a8-ad4b-78802712e7b0
https://suno.com/song/e61005b8-3eed-44f9-bf93-3cc1825d133a
https://suno.com/song/c6cc1ff2-f8e4-421a-963f-423561852544
https://suno.com/song/f52771dc-af08-40b7-a03d-71d52f166182

It seems to totally ignore the notes and rhythm, and only the word "triads" seems to have any effect.

This is on Windows, with the Brave browser, but the character symbols display and will be Unicode (most likely UTF-8 encoded) so platform shouldn't matter.

[sharp kraken]

im pretty sure the triads are all the same furry

the other notes played are different yes

but you can listen to the triads and here them exactly simillar in all cases

and yeah they all have triads as well

[charred merlin]

Not currently, I think they’ve mentioned “in painting” will be a feature of later versions so you can fix a specific spot.

As for right now you can create an extension point before the typo, or try to fix the error externally after separating the vocal stem.

[sharp kraken]

https://suno.com/song/a1d41e61-b2b2-4168-a1d9-963e29145c53

[solar rose]

Does anyone know what this "technique" is called when, for example, you go up a few semitones in pitch when repeating the chorus? Is that upshifting?

[sharp kraken]

https://suno.com/song/82e736f9-8813-4d0b-95b1-493fb5531376

[charred merlin]

Key change

[sharp kraken]

https://suno.com/song/90a692cb-ebd7-4b74-9a2c-81caafe664e2

alllllllll theeeeeeee saaaaaaaaaame

[sharp kraken]

[Soaring]

you can also use [final verse]

or combine them to get even more boost

key change is not quite gonna do that for you

[floral sonnet]

Ok next question, how do you get the AI to play specific pitches of notes, as in, rather than 12 not equal temperament but indian raga scale? is there a simple way or just experiment... i did get a weird thing yesterday when i prompted it with 'Micro-tonal Scale'

[sharp kraken]

dont quite understand the question like you mean C6 specific or what?

[floral sonnet]

...what we use in the west is called 12 note ET the a is set to 440 and the C2 is 261 hrz for example... but in indian music they have different notes....

[sharp kraken]

oh yeah. indian scales

japanese scales

only one it doesnt know is native american

hehehe ill have to teach it [comanche peyote scales]

im a peyote singer

but poor suno doesnt even know what a water drum is

[unique stone]

So - I may be totally getting this backwards as tbh my music theory is a bit lacking in many places 😅 so I'm trying to get an understanding of what exactly your example triads prompt is meant to do @sharp kraken - if I take the notes mentioned and play them as chords, I can't see how they relate to the outputs that are produced. What effect does the rhythm notation have?

[floral sonnet]

there is something called Just Intonation for example.. and loads of indian, balianese scales...

btw this is really cool --ᖶᕼᘿ ᘻᗩSᖶᘿᖇ ᘻᗩkᘿS ᗩ ᘺᗩᖻ

[sharp kraken]

im still working on understanding it all and ive told you as much as I can. If you don't think it works dont use it.

I think it works so i will continue to work it

[floral sonnet]

@charred merlin just finished scanning the WIKI .. but it is nowhere near complete... I wonder if a more complete master list of commands is to be found anywhere?... more on the technical music creation side?.. like the stuff we are talking about here...

[sharp kraken]

the timbre of an instrument can effect your perception more than tonality - so perhaps you are tricked by the timbre thats why you thinkits different

Ive made music for 32 years so I tune by ear, I can hear tonalpitch and seperate that from instrumental timbre

[sharp kraken]

theres a pretty good .pdf on that wiki

it has an extensive list of prompting commands you can use

but it doesnt tell you what they do

[sharp kraken]

This is my human brain music
https://soundcloud.com/joshuasca/feel-yur-pain?si=ab29cb29aa1b47c5bdebe44fcb69ae15&utm_source=clipboard&utm_medium=text&utm_campaign=social_sharing

[ornate belfry]

really sucks that sharing your music on suno shows your prompt. That slike a magician posting the secrets to his tricks

need to be able to hide that shit

[sharp kraken]

ask mr. tree

[floral sonnet]

@sharp kraken thanks'Feel ur pain'; ,,- listening!! ... where's the pdf ?.. cant ifnd it

[ornate belfry]

haha awesome

[sharp kraken]

https://www.suno.wiki/faq/style-and-lyrics/prompts-for-ai-dance-music/

[sharp kraken]

Hide prompts like this https://suno.com/song/b27de133-7b5d-45fc-b714-1c377d91b43f

[ornate belfry]

this is my fav song so far made in suno https://soundcloud.com/megachad9000/whos-in-the-house

[sharp kraken]

I always hope they read my prompts. I make it intresting like an alien book

full of ascii art

[ornate belfry]

lol thats actual magician magic

like sorcery really

[unique stone]

Being able to have this level of control over the music would be useful, which is why I wanted to see if it worked (along with using the trumpet and other instrument pictures.)

As generative AI can be quite creative and often not listen to prompts I know it will sometimes just do its own thing (e.g. I specify "pizzicato strings" and it might give me a piano instead!) I just wanted to understand how these kind of inputs are actually affecting the output because it would be really useful if they do, but my observations so far are that they don't (especially instrument/vocal gender icons.)

[sharp kraken]

100% real witchcraft i be doin

[sharp kraken]

well show me your prompts that didnt work, maybe your just making syntax errors

like - all your prompts might not work

have you ever prompted an entire instrumental?

[unique stone]

I went simple with this and just stuck a trumpet in the style prompt: https://suno.com/song/c5c8c5d6-b9de-4248-a5d6-05b2a16f2329