What actually works for duets in Suno

You asked for a duet. You got one singer performing both parts, or two singers who trade lines at random, or a man and a woman singing each other's words.

There is no duet button, and the workflows that people report succeeding with do not ask for one.

The practitioner's answer is to build it in passes

The most detailed public method comes from a songwriter with, by their own account, more than 300,000 streams, posted in March 2025. It opens with the flat version of the bad news: never try to get two voices inside one generation, because Suno does not work that way and you will not make it hold across a full song.

What they do instead:

1. Generate a 10 to 20 second stem with voice one, discarding anything with a flaw, spending one to two hundred credits to find a keeper.
2. Extend that stem two to four lines at a time until the first section is finished, for control over the phrasing.
3. Introduce voice two by changing something audible, with a vocal tag before the section and, critically, by starting the new audio immediately after the previous voice ends, which they place to the hundredth of a second in the edit window.
4. Alternate, using repeated lyrics to bring a known voice back.

Their aside is the most reusable line in the post: extend switches voices far more cleanly than replace section does.

The counter-claim comes with links

The top reply contradicts the headline flatly: it does work, and they got it twice in December, with two songs linked, one alternating a man and a woman line by line. The method they give is short. Say call and response in the prompt, name the voices you want, wrap the second person's lines in parentheses, and then press the button until it comes out right.

Both accounts can be true at once. One is a repeatable process and the other is a lottery ticket that paid out twice, and the disagreement is about how many generations you are willing to spend.

Syllable count is the lever people underuse

The same guide describes something more interesting than a tag: "if your chorus is sung slowly, with 4 syllables per line, but your verse is fast, with 10 syllables/line, suno will rap any 10-syllable additions, and sing any 4 syllable additions".

That is a delivery switch you write rather than request. Here are two parts from a song I wrote for this piece, written the way most people write a duet, with both voices at the same weight:

[Verse 2]
I have been counting the days since the letter came
I waited by the phone like a fool for an hour

Twelve syllables and 13, one after the other. Nothing there tells the model that the second line belongs to somebody else. Now the same exchange with the parts written at different lengths:

[Chorus]
You never call
You never call

[Verse 2]
and I am done pretending that I do not mind
I waited by the phone like a fool for an hour

Four syllables against 12 and 13. The short lines have room to be sung and the long ones do not, so they get delivered flatter and faster, and the two parts stop sounding like the same person.

Note

None of this is documented syntax. Suno publishes nothing about duet control, so every method here is somebody's working practice, dated, on the models they were using at the time.

What merges, and what people keep reporting

Two voices that are far apart in age seem to pull toward each other. On the big vocal descriptor thread from August 2025, somebody trying to write a musical with a child and an older man reports the two collapsing into one age: ask for the child and the man sounds young, ask for the older adult and the child grows up. Meta tags and the style prompt made no difference.

Another writer reports the opposite problem, a persona voice replaced by a new one between generations.

And the failure everyone recognises turns up as a joke. Under a parody post from May 2026, the reply people liked best singles out the part where the model goes out of its way to have the male and female singers swap each other's lines.

v6, launched on 9 September 2026, split the reports. Some users say duets work properly for the first time, with two voices singing together instead of trading lines mid-sentence, while others still cannot get v6 to follow directions between two named singers. Build in passes until the new model has a track record.

What none of this is

It is not a feature you can rely on. Every method above is a practice, reported by one person, on the model that existed when they wrote it.

It is not cheap. The workflow that works starts by spending 100 to 200 credits looking for a usable stem, before any of the interesting work begins.

And it is not a substitute for writing two distinct parts. If both voices say the same kind of thing at the same length, no tag, no persona and no timing trick will make them sound like two people.

Write your second voice's lines at a different length from the first, then build the song one voice at a time.