AI Music

Suno V6 Duet Prompts: Keep Each Singer on Their Own Lines

By Eddie Mathews·September 11, 2026·11 min read

The most common Suno duet failure is not a bad tag. It is that "duet" describes three different songs, and if you do not say which one you mean, the model picks. Get the structure right first and most of the voice-swapping problems stop being problems.

This guide is structure-first for a reason. V6 launched on 9 September 2026 and community reports on duet syntax genuinely contradict each other — some people got clean alternation on day one, others got voices trading places mid-line using what looks like the same approach. What holds up across those conflicting reports is the planning, not the tags.

Suno V6 duet prompts guide — three duet structures, mapping every line to a singer before generating, and setting Variety to zero so style tags survive

The short version

Name the structure first. Alternating verses, call and response, or harmonised chorus. They need different instructions.

Map every line to a voice before you generate. Ambiguity in the map becomes a swap in the output.

Put handovers on section boundaries. Mid-section switching is the risky move — confine it to one place.

Two roles is not two voices. Give each a distinct register, texture and placement.

Set Variety to 0 while debugging. Otherwise your style tags are being rewritten before generation.

Why duets fail more than solo vocals

A solo vocal asks the model to hold one identity for three minutes. A duet asks it to hold two, keep them distinct, and switch between them at moments you have specified only in prose. Every extra switch is another opportunity for the model to guess, and guesses are where swaps come from.

That framing is useful because it tells you where to spend effort. Reducing the number of ambiguous switch points does more for a duet than any amount of adjective tuning, and it is entirely within your control before you spend a single generation.

Structures to choose from

Map, written first

Variety, while testing

Risky switch region, max

Three duet structures, three different asks

Three Suno duet structures compared — alternating verses, call and response, and harmonised chorus, each needing different prompt language

Alternating verses — the reliable one

One voice owns a whole section, then hands over. This is the easiest structure to get right because the switch lands where the song already changes, so the model is not inventing a boundary. If duets have been failing for you, start here and add complexity only once it works.

Call and response — the hardest

Voices trade inside a section, often line by line. Every switch is mid-section, which means the model has to hold two identities simultaneously and alternate them on cue. This is where most reported swap failures live. It is worth doing — it is the most expressive of the three — but confine it to one section rather than spreading it through the song.

Harmonised chorus — not a switch at all

Both voices on the same words at the same time. This is a blend, not an alternation, and describing it with handover language is a reliable way to get neither. Say the voices sing it together and describe their relationship — who carries the melody, who sits underneath.

Build the singer / line map before you generate

This is the step that actually fixes duets, and it costs nothing because it happens in a text editor rather than in generations.

A singer and line map for a Suno duet, assigning each section to voice A, voice B or both, with delivery notes and the reason each handover lands where it does

Section	Voice	Delivery note	Why it lands here
Verse 1	A only	Close, conversational	Establishes A before anything trades
Pre-chorus	A only	Lifting	Clean runway into the first blend
Chorus 1	A + B together	Harmony, B below A	First time both are heard
Verse 2	B only	Warmer, further back	Handover lands on a section boundary
Bridge	A then B, line by line	Trading, no overlap	The one place call-and-response earns its risk
Final chorus	A + B together	Full harmony	Resolves to the blend the song built toward

Notice that only the bridge trades mid-section. Every other handover sits on a boundary the song already has. That is the whole trick: put your risk in one identifiable place, and let the arrangement do the work everywhere else.

Three lyric templates you can adapt

Write the assignment into the lyric itself rather than leaving it to a style-box description. These are shapes to adapt, not syntax with guaranteed behaviour.

Template 1 — clean alternation

Verse 1 all A. Chorus both. Verse 2 all B. Chorus both. Every switch is a section change. Use this to confirm the model can hold two identities at all before you try anything harder.

Template 2 — one trading section

Template 1, plus a bridge where A and B take alternate lines. Keep the lines short and complete so each one is a natural unit. If swapping appears, it will appear here, and you will know exactly where to look.

Template 3 — blend as the payoff

Verses alternate, first chorus is A alone, final chorus is both together in harmony. The blend arriving late gives the song a shape and gives you an easy A/B on whether harmony language worked.

In all three — label consistently

Use one labelling convention for the whole lyric. Switching between naming conventions mid-song is a documented way to confuse the assignment, and it costs you nothing to be consistent.

In all three — keep lines complete

A line that runs over a handover has no clean switch point. End the thought, then hand over. This single habit removes a large share of mid-sentence swaps.

In all three — Variety at 0

So the style description you wrote is the one the model reads. Debugging voice assignment while your prompt is being rewritten is not a debug at all.

Give each voice an identity, not a label

"Singer A" and "Singer B" tell the model there are two parts. They do not tell it the parts should sound different, which is why a technically correct duet can come back sounding like one person singing everything.

Dimension	Voice A example	Voice B example	Why it helps
Register	Higher, bright	Lower, warm	The clearest separator, audible instantly
Texture	Clean and close	Slight rasp, breathier	Distinguishes them even in unison
Placement	Front and dry	Half a step back	Makes harmony legible as two voices
Delivery	Clipped, rhythmic	Sustained, legato	Contrast survives even if timbre drifts

You do not need all four. Two strong contrasts beat four weak ones, and register plus delivery is usually the most robust pair, because they remain audible even when the model drifts on timbre.

Diagnosing a failed duet

Four symptoms, four different causes. Regenerating without changing the cause just re-rolls the same dice.

Four Suno duet failure modes and their causes — voices swapping mid-sentence, both voices sounding the same, one voice never appearing, and tags being rewritten by Variety

Symptom	Likely cause	What to change
Swaps mid-sentence	Ambiguous assignment at that line	Move the handover to a section boundary; end the line before switching
Both voices identical	Two roles described, one identity	Add register and delivery contrast, not more labels
One voice never appears	Structure mismatch — harmony language on an alternating map	Pick one structure per section and use only its vocabulary
Tags came back rewritten	Variety above 0 edited the style prompt	Set Variety to 0 while testing

What we cannot tell you yet

Two days of public V6 use is not enough to publish a syntax specification, and we are not going to pretend otherwise.

Launch-window reports conflict directly: within the same discussion threads you will find someone reporting a successful duet conversion and someone reporting that duets still fail, alongside a report of singers finally harmonising without swapping mid-sentence. Those are not all describing the same experience, which means no single tag recipe is currently defensible as reliable.

What survives that uncertainty is everything above: structure choice, line mapping, boundary placement, identity contrast and Variety at 0. Those are reasoning about how the request is posed rather than claims about undocumented behaviour, so they remain useful regardless of which tag convention turns out to win.

Releasing the finished duet

Once the voices are behaving and you have a track worth putting out, the creative problem is done and a separate one starts. Distributor AI screening runs on the exported file, once, at the moment you submit. It does not care how many singers are on the track or how good the arrangement is — it looks for the signal-level markers that identify AI audio. A flagged upload does not come back with a warning; it simply does not go live, so it never earns.

Undetectr handles that stage and works with Suno V6 output — its homepage leads with "Tested with Suno V6". It removes the AI watermarks and generation artifacts distributors screen for: embedded watermark signals, C2PA provenance metadata, vocoder spectral peaks and noise-floor patterns, so your track clears intake at Spotify, Apple Music and the rest and stays live. Around ninety seconds a track, output as MP3, WAV or FLAC, €39 one-time rather than a subscription.

Works with Suno V6

Two great vocals still meet the same scanner at upload.

Undetectr removes the AI watermarks and generation artifacts distributors screen for — embedded watermark signals, C2PA metadata, vocoder peaks and noise-floor patterns — so your V6 duet clears intake at Spotify, Apple Music and beyond. €39 one-time.

Clean your Suno V6 tracks →

Keep reading

Frequently asked questions

Eddie Mathews — AI Music Editor, Erasy

Eddie covers the craft side of AI music for Erasy — prompting, arrangement, vocal direction and the workflow between a generation you like and a file you can release. They test on their own material and say plainly where a technique stops working.

How do I make a duet in Suno V6?+

Decide the structure before you write a single tag, because 'duet' describes three different things and the model will pick one if you do not. Alternating verses gives each voice a whole section. Call and response trades line by line inside a section. A harmonised chorus puts both voices on the same words at once. Write your lyric so every line is unambiguously assigned, put your handovers on section boundaries wherever possible, give each voice a distinct identity rather than just a label, and set Variety to 0 so your style tags survive to generation.

Why do the singers keep swapping mid-sentence?+

Almost always because the assignment is ambiguous at that exact line and the model had no boundary to switch on, so it invented one. The fix is structural rather than a better tag: move the handover to a point where the song already changes — the end of a verse, the entry of a chorus — and the swap stops being arbitrary. If you genuinely need trading inside a section, confine it to one place such as a bridge, so there is a single risky region rather than a whole song of them.

Why do both my singers sound identical?+

Because you described two roles but only one instrument. Labelling them 'Singer A' and 'Singer B' tells the model there are two parts; it does not tell it they should sound different. Each voice needs its own register, texture and placement — one higher and closer, one warmer and further back, for example. Contrast is what makes alternation audible, and without it a technically correct duet sounds like one person singing the whole song.

Do bracket tags reliably control which singer sings?+

Treat them as something to test, not as commands with documented behaviour. V6 launched on 9 September 2026 and community reports on duet syntax genuinely conflict — some people report clean alternation on day one, others report the same approach failing. Suno has not published a duet tag specification. So use the tags that appear to work for your material, verify on your own song, and be sceptical of any guide presenting a syntax as deterministic this early.

Is it better to generate a duet from scratch or convert an existing song?+

Both routes have launch-day reports of success, which is why the honest answer is to try both on your own material. A fresh generation gives the model the whole arrangement to build around two voices from the start. A conversion workflow starts from something whose structure you already like and changes the vocal layer, which can preserve an arrangement you do not want to lose. If one route keeps producing swaps, switching route is a cheaper experiment than another twenty regenerations.

How do I get actual harmony rather than alternation?+

Ask for it explicitly and stop using alternation language anywhere near that section. Harmony is not a switch between voices, it is both voices on the same words simultaneously, and mixing the two vocabularies in one request is a common reason people get neither. Say that both voices sing the chorus together, and describe the relationship between them — one above the other, one carrying the melody while the other supports. Then keep the alternation instructions strictly in the sections that alternate.

Does Variety affect duet prompts?+

Yes, and it is worth setting to 0 while you are working this out. Suno documents the Variety slider as adjusting and updating your style prompts before generation, so raised Variety means the duet description reaching the model may not be the one you wrote. When you are debugging voice assignment, that is an unnecessary extra variable. Set it to 0 to retain control of your style tags, get the structure working, then raise it later if you want the model to explore around a brief you already trust.

Will a duet track pass distributor screening?+

The number of singers makes no difference to screening, which looks at signal-level markers in the exported audio rather than at musical content. What matters is that the file is cleaned before submission, the same as any other AI-generated track. Screening runs once, at upload, and a flagged track does not go live and therefore never earns. Run the export through watermark and artifact removal before you send it to Spotify, Apple Music or any other DSP.

Disclosure: Erasy is an independent guide to AI music cleanup, and Undetectr is the tool we recommend and link to. V6 launched on 9 September 2026 and community reports on duet syntax conflict, so the structures and templates here are reasoning about how to pose the request rather than claims about documented model behaviour. Test tags on your own material before relying on them. Figures current as of 11 September 2026.