Turn lyrics into a song: an AI song generator from lyrics

A finished lyric is one third of the job

A friend sent me 32 lines she had worked on for two weeks. Real lines, the kind with a specific object in them: a bus ticket left in a coat pocket, not "my broken heart". She pasted them into Suno, left the Style box empty, pressed create, and got back a mid-tempo piano ballad with a polite female voice that sounded like the hold music at a dental clinic. The lyric was not the problem. The empty box next to it was.

Suno reads three inputs. Your words decide what gets sung. The Style line decides what the band plays and how the singer sings it. Skip the second one and the model fills the gap with its own average, and its average is the safest pop arrangement it knows how to build. That is the whole reason a good lyric pasted into an otherwise empty form comes back sounding like everybody else's, and it is why the fix is almost never "rewrite the lyric".

The three fields, and what belongs in each

Turn on Custom mode first. The simple mode gives you one box and guesses the rest, which is fine for a joke and useless for a song you care about. Custom splits the form into parts you can actually aim.

Lyrics. The sung text plus the structure tags. Nothing else lives here. No notes to yourself, no "(guitar solo)", no alternate line you were still deciding between, no translation in parentheses.

Style (labelled Style of Music in some places, just Style in others). One short English line of descriptors: genre, tempo feel, instrumentation, vocal type, production character. melodic techno, 124 feel, analog bass, breathy female vocal, wide reverb, cold steers harder than three sentences of adjectives. Write it in English even when the song is in Russian or Spanish, because the words in Style are instructions to the model and only the words in Lyrics are a performance. A fuller treatment of what belongs on that line sits in the Suno prompt guide.

Title, and Exclude Styles. The title box is cosmetic. It names the file and nothing else. Exclude Styles is the opposite of cosmetic: every "no" goes there, and it is the field most people never open. Suno treats a negative inside Style as a topic rather than a prohibition, so no autotune sitting in the Style box regularly produces autotune. Move the bare word autotune into Exclude and the ban holds. Same trick for male vocal, distorted guitar, trap hi-hats, spoken word.

The prompt generator here builds that Style line and that Exclude list for you, and it is worth saying plainly before we go further: nothing on this site produces audio. Suno performs. We write the paperwork.

The rule almost everyone breaks

Everything in the Lyrics field that is not inside square brackets gets sung.

Not interpreted. Not read as context. Sung, out loud, in the voice you asked for, in the middle of your song. Paste this:

Verse 1
(soft, almost whispered)
I found the ticket in your coat

and the vocal opens with the words "verse one, soft, almost whispered" before it ever reaches the ticket. I watched this happen to a wedding song at two in the morning, and the singer announced "chorus" in bar nine, directly before the bride's name.

Square brackets are the only escape hatch. [Verse], [Chorus], [Bridge], [Intro], [Outro], [Pre-Chorus] shape the arrangement and never reach the microphone. You can stack instructions inside one bracket with pipes: [Outro | fade out | instrumental]. Anything the model should know but no listener should hear has to live inside brackets. Which tags genuinely change the output and which are decorative superstition is catalogued in the metatag rundown.

So converting a lyric document into a Suno lyric is mostly deletion, and that surprises people who arrived expecting to add something.

Formatting a lyric you wrote somewhere else

Strip the parts that were never meant to be heard

Open the draft and delete the chord symbols, the V1 / C / V2 shorthand, the parenthetical performance notes, the alternate line you left in as a maybe, the working title at the top of the page, your own name. Then put structure back as bracket tags.

Before, straight out of a notes app:

UNTITLED (draft 3)
Am          F
V1 (quiet, close mic)
I found the ticket in your coat   / or "jacket"?

After:

[Verse]
I found the ticket in your coat

Four lines became two, and only one of them will be sung. That ratio is normal.

A line break is a breath

Suno lands one melodic phrase per line, so line breaks decide where the singer inhales. An eighteen-word line gets crammed into a bar that cannot hold it, and the delivery either accelerates into gibberish or the tail gets clipped. Break anything past roughly twelve syllables into two lines. Going the other way, if you want two short phrases sung as one continuous run with no gap, put them on the same line and let the comma do the work.

One blank line between sections. No blank lines inside a verse, unless you actually want a hole in the middle of it, because the model reads that gap as a structural break and will happily drop four bars of piano into it.

The spoken intro

This is the one place I will not promise you a result. Suno wants to sing. Tag the line [Spoken Intro] or [Spoken Word], keep it to a single sentence, and add spoken word intro to the Style line so both fields agree. On a good run you get speech. On a bad run you get the same sentence sung, at which point the honest move is to cut the intro and start on the hook rather than burn eight regenerations chasing it.

Repeated choruses

Write the chorus out in full every single time it appears. Do not write [Chorus x2], and never write (repeat chorus), because that one is not in brackets and the singer will happily perform the words "repeat chorus" for you.

An empty second [Chorus] tag does not reliably pull the earlier text back. Writing the words out again also buys the thing you actually want: when the words are identical, Suno usually reuses the melodic shape, and a repeating melody on repeating words is the definition of a chorus. If you want a final-chorus variation, change one line and leave the rest byte-for-byte identical.

Length: your lyric is probably too long

The most common complaint from people pasting real lyrics is that the track ends mid-sentence. The cause is arithmetic. A generation covers a fixed span, and how long that span runs depends on the model version and the plan you are on, so check the current numbers instead of trusting a figure from a forum post; the practical side of that is worked through in the full-length song guide.

The useful proxy is line count. A classic verse, chorus, verse, chorus, bridge, chorus at four lines a section is 24 lines, and that sits comfortably. Sixty lines does not, no matter how good they are. When you overrun you get one of three symptoms: the track stops mid-lyric, the vocal accelerates into a patter to fit everything in, or the bridge quietly disappears.

The fix is cutting the lyric. Not asking for a slower tempo, which makes the model compress harder. Cut in this order: the third verse, the pre-chorus if the chorus already lands on its own, any adjective inside a line that a listener would not miss, and the second of two verses making the same point. Between two verses saying the same thing, keep the one with the concrete noun in it.

The route I would actually take on this site

Two tools, in this order, and neither one makes a sound.

Step one applies only if you do not already have words. The song generator takes a brief and hands back a title plus a complete lyric already split into [Verse], [Chorus] and [Bridge]. The brief is a fixed set of slots: a theme up to 300 characters, point of view, one of nine song languages, a target audience age range, a structure (classic, no-bridge, double-chorus, story, loop, ballad), a hook direction, a vocal setting, one of 53 root styles, and up to three emotional tones. It runs 20 credits, roughly $0.20 at 100 credits to the dollar, though the pricing page is the source of truth for what that costs today.

Step two applies always. Run the prompt generator for the Style line and the Exclude tags. It is not a model, it is deterministic: the same slots return the same recipe every time, and every "don't" you tick gets flipped into a positive tag in Exclude. That predictability is the entire point when you regenerate. If the recipe wobbles between runs, you cannot tell whether the change you hear came from the one slot you moved or from the tool being moody.

Step three, paste. Lyrics into Lyrics, Style line into Style, Exclude tags into Exclude Styles. Suno does the performing.

Anyone selling you an AI song generator from lyrics that returns a finished master is either describing Suno or wrapping something like it. There is no third category yet.

When it goes wrong

Suno sings your stage notes. All of them are outside brackets. Bracket them or bin them.

Two verses merged into one long section. Either the blank line between them is missing, or the tag is sitting on the same line as the first lyric. [Verse] needs a line of its own.

The chorus never repeats, or arrives different every time. The text is not identical. Copy and paste it instead of retyping, then shorten it to four lines if the melody still drifts.

The track ends mid-lyric. Too many lines. Cut, do not slow down.

The voice is wrong. That is not a Lyrics problem and no amount of editing the words will fix it. The voice lives in Style, and the negative half of it lives in Exclude. Editing passes on the words themselves are covered in the lyrics editing guide.

FAQ

Can I paste a poem straight in? You can, and it usually sags. Poems have no chorus and no repetition, so the model gets nothing to build a structure on and defaults to through-composed drift. Pick the two strongest lines, repeat them as a [Chorus], and split the rest into verses.

Does the lyric language have to match the Style line? No, and it should not. Style stays English because Suno parses English descriptors far more reliably. The sung words follow whatever language the song is in.

Does the Title field affect the music? Practically no. Treat it as a filename. The work happens in Style and Exclude.

What if I only have a chorus? Then the hardest part is done. Write the chorus twice with a [Verse] tag between them, generate, and listen to what the arrangement suggests for the empty verse. Or feed the chorus theme into the song generator and let it build verses around it.

Can I use lyrics I did not write? Someone else's lyric is someone else's copyright, and generating a track from it does not launder that. The commercial side is unpleasant and worth reading before you upload anything: what you can actually sell.

Will Suno fix a weak rhyme? No. It sings what you give it, weak rhyme and all. A near-rhyme that reads badly on the page often sings fine, though, so test before you rewrite.