What does Suno do to your lyrics, and how do you write so it leaves them alone?

The stage direction that got sung

Paste this into the Lyrics field and hit generate:

(whispered, barely audible)
I left the keys under the mat
(chorus, loud)
and I am not coming back

Back comes a singer delivering "whispered barely audible" with real feeling, on a melody, then handing you "chorus loud" with the same commitment, as though it were the second line of the verse. The round brackets protected nothing. To the model they are letters, no different from the rest.

One rule governs that field and it does not bend: everything not enclosed in square brackets gets sung. Round brackets hide nothing. Neither do asterisks, all caps, a row of dots, or the word NOTE parked in front of the line. If you typed it and there are no [ ] around it, a voice will find a melody for it.

The mistake is cheap and it sits on the surface. What happens to the lines you actually meant to sing is the more interesting part. The model re-reads the sheet, splits it into sections of its own choosing, decides which block behaves like a chorus and where a line will not fit the bar, and on the way it makes decisions you believed were already made. Most of the disappointment people file under "bad AI vocals" grows out of formatting the writer handed away without noticing.

Your sheet works as a request, not as a score

Write four stanzas of roughly equal size with no tags anywhere. You are certain you wrote verse, chorus, verse, chorus, because that is what you had in mind. The intention never arrives. What arrives is blocks of text separated by blank lines, and the model picks which block behaves like a chorus.

It picks on surface evidence: what repeats, what is shorter, where the line endings rhyme more tightly, what comes back after other material has intervened. If your chorus happens to be the longest and least repetitive chunk on the page, expect to hear it down in the low narrative part of the melody while verse two climbs instead. Nothing malfunctioned. The song is coherent. It is just a different song.

Easier to show. Two sheets, one story, roughly the same length.

Sheet A, no markup:

The hallway light goes out a little after seven
I count the floors the way that everybody counts them
You said that you would write to me by morning
I read it twice and put it back in my pocket

The hallway light comes on a little after seven
I count the floors and nobody comes down them
You said that you would call my name by morning
I said it once myself and left the city

Two blocks of one size, built from the same turns of phrase, which leaves the model no handle for naming either one the chorus. It will name one anyway, and whether that matches your plan comes down to luck.

Sheet B, the same material, cut down and marked up:

[Verse]
The hallway light goes out a little after seven
I count the floors the way that everybody counts them

[Chorus]
Put it back
Put it back
You said that you would write to me by morning
Put it back

[Verse]
The hallway light comes on a little after seven
I count the floors and nobody comes down them

Now the chorus lines are visibly shorter than the lines around them, Put it back returns three times inside it, and the morning line is the one thing in the block that does not come back, which is exactly why it lands. Sheet B is not better writing than sheet A. It is writing built to survive being read by a program that has no idea what you meant.

The full catalogue of shapes worth marking up is laid out in the piece on song structure templates. This article works the layer underneath: what the text itself does once the model has it.

What we actually sell, so nobody has to guess

Before going further, the part usually buried at the bottom of the page.

We write words. That is the whole product. The song lyrics generator takes a theme of up to 300 characters plus a set of closed lists: point of view, song language, audience age, structure, hook direction, vocal type, one of 53 curated musical styles, and between one and three emotions out of twelve. It returns a title and a finished lyric of up to 20,000 characters. It costs 20 credits, which is $0.20. Duet is not among the three vocal options, so if you want two characters trading lines, that structure is something you build by hand.

We do not produce sound. No vocals, no instruments, no mp3, not one second of audio. To hear the words, the sheet has to travel somewhere that draws music. Usually that is Suno, and Suno is a separate company. We do not own it, do not resell it, do not steer its behaviour, and we find out about changes to it the same day you do. Everything below describes how text behaves under a model reading it, which holds up better than the position of any button in any given month.

That split is why the advice here still pays off for people who never open Suno at all. A sheet that reads cleanly for a machine reads cleanly for a session singer, and for the same reason: the structure is on the paper instead of in your head.

A long line breaks, and not where you would have broken it

A line of lyric has to fit inside a musical phrase. When it runs longer than the phrase, something gives, and the model gives in one of three ways: it compresses the syllables into a patter, it stretches the last word into a held note to fill out the bar, or it breaks the line and pushes the tail into the next phrase.

The break is the dangerous one, because it does not happen at your comma. It happens where the melody ran out. Look:

I remembered your number on the bus on the way to the airport and I never called

Twenty two syllables in a slot asking for nine or ten. The tail, "and I never called", carries the entire meaning, and the tail is precisely what gets shoved into the following phrase, where it arrives late, with no accented landing spot, after the listener has already moved on. You wrote a punch. It came out as muttering after the fact.

Split it into two lines and the punch gets its own runway:

I remembered your number on the bus
and I never called

Ten syllables and five. Notice what got fixed besides length. The second line is now noticeably shorter than everything around it, and that reads as an author's decision rather than a stump. Asymmetry you chose sounds like craft. Asymmetry the model imposed sounds like a defect.

The working range for a sung line in most popular forms sits somewhere between five and twelve syllables, with the long end belonging to rap and the short end to chants and hooks. That is not a rule handed down by any service. It is just how much text fits into four beats at a walking tempo. Count on your fingers. It takes eleven seconds and saves a generation.

A chorus needs repetition more than you do

Poets get trained out of repeating themselves, and here that training works directly against you.

Repetition inside a chorus does three jobs for the model at once. It flags the block as a chorus, and the arrangement lifts. It hands the melody a phrase that can be reused, which is the whole difference between a hook and a fourth verse. And it gives the vocal, human or synthetic, something to push off, because ornament goes on the second pass, not the first.

Compare two choruses built from roughly the same amount of material.

With variety:

Every window on this street is dark tonight
The whole avenue gave up on waiting for the morning
Not a single lamp is burning in the building

With repetition:

Every window dark
Every window dark
Every window dark tonight
Leave a lamp on for somebody

The first one is better prose. The second comes back sounding like a song, because the return of Every window dark is the peg the melody hangs on, and the lamp line ends up the only surprise in the block. Songwriters put the knife in the last line of a repeating chorus for a reason.

If your choruses keep coming out hollow, the craft side of the problem gets taken apart in how to write a chorus.

This advice has a ceiling, and it is worth naming before it turns to mush. Four identical lines with nothing deviating give the model no material to shape, and instead of a chorus you get a football chant. The boundary sits roughly where two or three repetitions sit next to one line doing something of its own.

Numbers, abbreviations, rare words, and anything a reader would have to look up

Any oddity in the text is a place where pronunciation becomes a guess, and a guess sung confidently sounds worse than an ordinary word sung ordinarily.

Numbers. Spell them, do not type digits. 1985 can arrive as "nineteen eighty five", as "one thousand nine hundred and eighty five", or as a mumbled run in which the syllables you counted evaporate. Write "eighty five" and the syllable count is locked before the model ever opens the file. Same for 3 a.m. against "three in the morning", and same again for prices, ages, and house numbers.

Abbreviations. Some get read as words, some get spelled out letter by letter, and the capitals alone do not tell you which way a given one will go. Both readings are legitimate English. Only one of them fits your bar. If the letters matter, space them out. If the sound matters, write the sound.

Proper nouns and rare words. An unfamiliar surname, an invented place, a technical term: every one of them puts the stress placement on a coin toss. You can spend the coin deliberately, or you can write the word in the line the way it sounds and leave the correct spelling to the title, where nothing gets sung.

Words carried in from another language. A foreign word dropped into an English line reliably produces an odd accent in the middle of a phrase. Sometimes that is the effect you want, so keep it. More often it is a leftover from the draft. Decide on purpose, and if the song is meant to be bilingual, the trade-offs are worked through in the piece on writing songs in another language.

The punctuation you are counting on. An ellipsis does not become a pause. Neither does a semicolon. If you need a breath, put in a line break, which works far more often than a comma ever does.

Old poetic spellings. blessed is one syllable as a verb and two as the adjective the hymn-writers used, and the page looks identical either way. If the count depends on it, pick a different word rather than betting on a diacritic nobody renders.

A blank line is an instruction, whether or not you meant it that way

The blank line is the quietest markup in the field and one of the loudest in the result. It draws a boundary and announces that one piece has ended and the next begins.

Which means a stray blank line in the middle of a verse cuts it into two pieces the arrangement may treat differently. And a missing blank line between verse and chorus glues them into one long block that the model handles as a single unit, dissolving the contrast you wrote them for.

Three habits cover nearly all of it:

• One blank line between sections. Not two, not zero.
• No blank lines inside a section, even when your eye wants air. Your eye wants air. The parser sees a wall.
• No blank lines at the end of the sheet and no spaces hanging after the last word. They are invisible on screen and they are real characters in the field.

The most common version of this breakage looks like this. Somebody writes in a word processor with generous paragraph spacing, copies the whole thing, and every stanza arrives pre-split into couplets. The result sounds chopped, the writer blames the model, and the culprit was the space-after-paragraph setting in the editor.

Stress: the thing that sounds wrong when nobody can say why

This is the one people describe as "something is off and I could not tell you what".

Every word has a stressed syllable and every melody has strong and weak beats. When a stressed syllable lands on a weak beat, or an unstressed one gets handed a big note, listeners hear the fault instantly and usually cannot point at it. In your own lyric you will not notice, because you read it with the stresses you intended rather than the ones the words actually carry.

English hides this differently than people expect. Stress here is mostly fixed per word, so the trap is not that a word wanders. The trap is a family of words that changes stress when it changes job, with the spelling staying put: record, present, object, contract, permit, rebel. The noun takes the first syllable, the verb takes the second. In prose the surrounding grammar settles it. Lyric lines throw the grammar away, which is when you find out who is deciding.

The sharpest case is refuse, where the two readings differ in stress and in the final consonant. As a noun it is REF-use, rubbish by the door. As a verb it is re-FUSE, to turn something down. One spelling, two words. Put it at the end of a line, where the rhyme hangs on it, and the choice walks out of the building with your rhyme:

I keep it by the door with all the rest, refuse

A model reading the line without leaning on grammar can take either one, and neither is yours any more. Move the word out of the final slot and give it a subject, and the ambiguity disappears:

I refuse it at the door and keep the rest

I refuse can only be the verb, and the line now ends on rest, one syllable, one pronunciation, nothing left to decide. Small move. It is the move that separates a line that sounds sung from a line that sounds read out.

Compound stress is the same trap in a different coat. A greenhouse takes the stress up front; a green house puts it on the second word. The page cannot tell you which one the model will build, and the beat does not care about your intention.

How to check stress before you spend a generation

Two checks, neither requiring software.

Read the line out loud flat. Not with expression. Deliberately flat, tapping an even pulse on the table with your hand. Expression smuggles your intended stress back in and hides the problem. Flat reading against a pulse exposes it. Any syllable that argues with the tap needs to move.

Move the word, not the melody. When a stress lands wrong, the fix is the line, not the arrangement: find a shorter synonym, reorder the phrase, or pull the disputed word off the end of the line into the middle, where a weak landing costs nothing. Line endings are expensive real estate. Only words whose stress you are sure of should live there.

The rhyme side of the same question, which endings genuinely close and which only pretend to, gets worked out in rhyme schemes for songwriters.

Lyrics for a machine and lyrics for a live singer are built differently

A session singer reads your sheet and fills in what is missing. Sees a weak line and phrases around it, sees a stress collision and slides past it, hears the word "bitter" and adds bitterness without being asked. Thirty years of listening to songs tells them where the accent goes, and they throw that in for free.

The model brings the average of everything it has read. Where the singer interprets, it fills the gap with the mean. Four consequences follow for how you write.

Nothing is implied. The singer knows the second chorus should be bigger. Write it bigger, mark it up, or accept a flat one.

Ambiguity resolves against you, toward the ordinary. If a line can be read tender or barbed, expect tender, because tender is simply more common in the songs the model learned from. If you want venom, put it in the words rather than the delivery.

Consistency buys more than variety. A singer welcomes a surprising line. The model treats surprise as noise and tends to smooth it. So one large departure belongs in a marked place, the last line of a chorus or the bridge, rather than scattered as small deviations across every verse, where they get averaged to nothing.

You are the arranger now. Length, repetition, section boundaries and stress stopped being matters of taste, because you have no other channel. Anything not encoded in the text gets decided by something with no stake in your song.

None of that means writing worse. It means the decisions have to be visible on the page. Good songwriters already work this way, because a sheet handed to a live band runs into exactly the same problem.

Editing the sheet before you paste it

Ten minutes down this list saves more generations than any amount of tag hunting.

1. Strip every note to yourself. Everything in round brackets, everything in caps, any line that talks about the song instead of being the song. If it needs to survive as an instruction, move it into square brackets. If it does not, delete it.
2. Count the syllables in every line. Write the numbers in the margin. A line more than two or three syllables away from its neighbours inside the same section will be compressed, stretched, or broken.
3. Check the section boundaries. One blank line between, none inside, no trailing spaces.
4. Make the chorus a chorus by repeating something. At least one phrase has to return inside the block, because a block where nothing returns behaves as a verse no matter what you labelled it.
5. Read one column: only the last word of each line. Top to bottom. That is where the rhyme lives, where the stress risk concentrates, and where the listener's attention goes.
6. Spell out every number, every abbreviation, every rare name. Or leave the coin spinning, knowingly.
7. Read it aloud flat against a tapped pulse. Mark every syllable that fights the tap.
8. Cut one line. Not because it is bad. The line you end up cutting is almost always the one that was padding out the bar, and the sheet tightens the moment it leaves.

The list works on any sheet, including the one you wrote yourself at three in the morning. If starting from a blank field holds no appeal and those ten minutes are better spent editing a draft that already exists, build a sheet in the generator from your own theme and style, then run the result down the points above. The generator hands you a structured draft for $0.20. It becomes yours during the edit. The credit arithmetic lives on the pricing page.

What we cannot promise

Suno belongs to Suno. It changes its behaviour, it changes its interface, and we hear about it when you do. Any paragraph resting on a specific screen or a specific setting would be stale within a month, so beyond the name of the lyrics field there are no paragraphs like that here. What is here is a description of how a model reads a page of text, and that has held long enough to build a workflow on.

Two things are worth keeping apart in your head. The bracket rule is mechanical and dependable: anything outside square brackets gets sung. Everything else on this page describes a tendency, not a guarantee. The model can surprise you, the surprise is sometimes better than your plan, and the right response in that case is to keep it.

And the honest version of the division of labour: we write words for twenty cents and we make no sound at all. If you need a finished track, the words are step one and the singing happens somewhere else. The road from a sheet to something you can press play on is mapped out in how to turn lyrics into a song. And if the place you are stuck is the writing rather than the formatting, start with how to write song lyrics.