Suno Duets: How to Stop the Singers Swapping Mid-Line

Three things people mean by "duet"

Type "duet" into a Suno style prompt and you are asking for something the model has to guess at, because a Suno duet is really three different structures and they behave very differently.

Alternating verses. Voice A takes verse one, voice B takes verse two, both take the chorus. This is the classic pop duet shape and it is by far the most achievable.

Call and response. The voices trade within a section, often line by line or phrase by phrase. Harder, because the switching boundaries are much tighter.

Simultaneous harmony. Both voices sing at once, in harmony rather than in sequence. This is the hardest and least reliable, because it asks for two rendered voices in the same moment rather than one after the other.

Most people asking for "a duet" picture the first, describe none of them, and get an unpredictable blend of all three. That unpredictability is the mid-line swapping everyone complains about.

Assign lines in the lyrics, not just in the style box

This is the single change that moves duet results the most.

A style prompt describing "a male and female duet" tells the model what the track should sound like in general. It does not tell the model who sings line four. Voice then behaves like a texture that can drift, because nothing in the input marks where one singer stops and the other starts.

Section labels in the lyric sheet give it those boundaries. The pattern that works looks like this:

[Verse 1 - male vocal, low and restrained]
...lyric lines...

[Verse 2 - female vocal, brighter and higher]
...lyric lines...

[Chorus - both voices together]
...lyric lines...

Three things are doing work there, and all three matter.

The label names the section, which Suno already uses for structure. The label names the voice, which converts singer from texture to role. And the label carries a short descriptor, which keeps the two voices from converging on the same sound.

That last point is worth dwelling on. Two voices described in similar terms tend to blend. Two voices with genuine contrast in register, texture and delivery stay separable. If your duet keeps collapsing into one singer, the fix is often not more instruction but more difference between the two descriptions.

Templates for each structure

Alternating verses

The most reliable shape. Keep the switches on section boundaries, where the model is already making a transition.

[Verse 1 - male vocal, warm baritone, conversational]
[Pre-Chorus - male vocal continues]
[Chorus - both voices, female on the melody, male harmonising below]
[Verse 2 - female vocal, clear mid-range, more urgent]
[Chorus - both voices]

Notice the chorus specifies who does what rather than just "both". "Both voices" invites a blend; naming the melody and the harmony line gives each voice a job.

Call and response

Harder, because switches happen inside a section. Keep the exchanges short and regular rather than irregular.

[Verse 1 - call and response]
(male) line one
(female) line two
(male) line three
(female) line four

The inline parenthetical carries the assignment where a section label cannot reach. Expect this to be less reliable than alternating verses, and expect it to work better in slower material where the phrase boundaries are wide.

Simultaneous harmony

Ask for it explicitly and keep it contained.

[Chorus - both voices in harmony, sung together, female lead with male third below]

Two rules of thumb. Specify the harmonic relationship rather than saying "harmony", and keep the harmonised sections short. A whole track of simultaneous two-voice rendering is asking a great deal; a harmonised chorus against alternating verses is much more achievable.

Two settings that change duet behaviour

Before you conclude a duet prompt does not work, check two controls, because both are documented to affect exactly this.

Set Variety to 0 while you are debugging. Suno's v6 FAQ states that the Variety slider introduces variety by adjusting and updating your style prompts, and that reducing it to 0 retains full control of your style tags. If your carefully contrasted voice descriptions are being rewritten before generation, you are not testing the prompt you wrote. Get a clean baseline, then raise Variety again later if you want more variation. Our guide to the sliders covers the rest of the controls.

Try Max Mode on the duet. Suno documents Max Mode as being best for, among other things, "keeping vocals and style consistent through the whole track." Vocal inconsistency across a track is the precise complaint duets generate. It costs more credits, so run it against a standard generation of the same prompt rather than assuming it helps.

What v6 actually changed for duets

Honestly: the evidence conflicts, and it is only days old.

Some users report clean results in v6, including successful conversions of existing tracks into duets. Others report duets still failing in the same ways they failed before. At least one report describes singers finally harmonising without swapping mid-sentence, which would be a genuine improvement if it holds.

That spread is what you would expect two days after a model launch, when everyone is testing different material with different settings and different prompt habits. Our v6 review covers the wider reaction, including the broader vocal-phrasing complaints that may be related.

Treat the threads as hypotheses, not findings. The useful move is to test one structure with one fixed lyric sheet, change one thing at a time, and build your own view. That is genuinely faster than reading another forty comments.

A debugging order that saves credits

When a duet comes back wrong, work through this in order rather than changing everything at once.

One: Variety to 0. Otherwise you are debugging a rewritten prompt.

Two: add voice labels to every section. If your lyric sheet has bare [Verse 1] labels, that is the highest-value fix available and it costs nothing.

Three: increase the contrast between voice descriptions. If both singers are described as "smooth" and "melodic", make one gravelly and low and the other bright and high. Blending is usually a contrast problem.

Four: move switches onto section boundaries. If you are attempting call-and-response and it is failing, try the same lyric as alternating verses. If that works, the structure was the problem, not the prompt.

Five: try Max Mode. Documented for vocal consistency, which is the failure you are chasing.

Six: accept the ceiling. Simultaneous two-voice harmony across a full track is at the edge of what the model does reliably. If four attempts have failed, restructuring the arrangement will get you a finished song faster than a fifth attempt at the same ask.

The part that is not about prompting

Whichever structure you land on, the arrangement has no bearing on whether the track gets released.

Distributor screening at ingestion reads the artifact layer: the spectral fingerprints, phase relationships and micro-timing distributions that are byproducts of how any AI model synthesises audio. A duet carries them exactly as a solo vocal does. There is no vocal arrangement that clears DistroKid and no harmony that satisfies a classifier.

That step happens once, on the exported file, before you submit. Our artifact removal ranking covers the tested options across a 50-track, six-distributor corpus.

The short version

Decide which of the three duet structures you actually want before you write a word of the prompt. Alternating verses is the reliable one; simultaneous harmony is the hard one.

Assign voices in the lyric sheet with section labels, not only in the style box, and make the two voice descriptions genuinely different from each other. Most mid-line swapping is a missing-label problem, and most blending is a contrast problem.

Set Variety to 0 while you debug so you are testing your own prompt, and try Max Mode, which Suno documents for exactly the vocal-consistency failure duets produce.

And keep the release step separate. How the voices are arranged decides how the song sounds. Whether it reaches Spotify is decided by something the arrangement cannot touch.