How do you create an AI music artist with a consistent voice and style?

The first single from a made-up act called Lowtide Mira sounds right: a breathy alto over warm synths, a chorus that sits low and sad. The second single, made a month later with "the same prompt", comes back with a brighter, younger singer and a drum machine that wants to be house music. By the third release the comments ask whether this is a playlist of different artists. Nobody did anything wrong in any single session. The act fell apart in the gaps between sessions, because nothing was written down and nothing was locked.

That is the real job behind "create an AI artist". One good generation is easy. The tenth generation sounding like the same person as the first is the work, and it is mostly bookkeeping, not inspiration.

An AI artist is four things that each drift on their own

People talk about "the AI artist" as if it were one setting. It is four separate layers, each controlled by a different tool, and each one drifts if you leave it to chance.

Notice that only one row is about Suno itself. The rest are decisions you make once, write down, and then refuse to renegotiate every time you open a new session.

If you are still deciding whether an AI act can earn anything at all, read what actually makes money with AI music first. That page is for the reader weighing the business. This one is for the reader who has decided to build the act and needs it to hold together across a year of releases.

Step 1: write the persona sheet before you generate anything

Every consistent AI artist has a document behind it, even if it is a note on a phone. Without it, each session starts from memory, and memory rounds everything off. Here is a complete sheet for the invented act from the opening. Copy the shape, not the content.

ARTIST        Lowtide Mira (invented act, solo, female vocal)
ONE LINE      Late-night songs about leaving coastal towns and not quite leaving

VOICE         breathy alto, close-mic, restrained, never belted
VOICE SOURCE  Suno Persona "Mira v1", made from release 01, set to private

CORE STYLE    dream pop, breathy female alto, warm analog synths,
              vintage drum machine, reverb-washed guitars, melancholic, mid-tempo
EXCLUDE       rap, screaming, autotune, EDM drop
MAY VARY      one texture per song: tape hiss | piano | strings | acoustic guitar
NEVER         genre swap, choir, upbeat major-key anthem

LYRIC BRIEF   POV first | emotion longing, loneliness, calm | language en
              structure classic or ballad | hook title-line or confession
IMAGE BANK    tide charts, bus stations, salt on windows, borrowed coats
BANNED WORDS  fire, soul, forever, heart of gold

CADENCE       one single every 4 weeks, EP after 4 singles
DISCLOSURE    AI-assisted, stated in bio and in credits where the distributor allows

Two lines in that sheet do most of the work. MAY VARY gives each song one controlled difference, so the catalogue does not become ten copies of one track. NEVER names the moves that would break the act, so you stop yourself before a "just this once" experiment ships under the artist's name. Experiments are fine. They go out under a different name.

The image bank and the banned words matter more than they look. Listeners recognise a songwriter by recurring objects and by what that writer never says. A banned-words line is the cheapest consistency tool in this whole article.

Step 2: lock the voice, and know which Suno tool you are actually using

Suno has three different features that people lump together as "keeping the voice". They solve different problems. All details below are from Suno's own help and release pages at the time of writing, and plan access changes, so check Suno's pricing page before you plan around one.

Persona. You open a song you already generated, choose Create, then Make Persona, and Suno saves the feel of that song, its vocals and style, for reuse. When you select the Persona in Custom mode, its style details fill the style field automatically. Two consequences for an artist project. First, the source song is everything: a Persona made from a muddy take carries the mud into every release. Second, Suno's help page says Personas are public by default. If the voice is your act's identity, switch it to private, or someone else can sing their songs with your artist's voice.

Voices. This is the feature for using your own recorded or uploaded voice. Suno's Voices page describes a verification step meant to keep an uploaded voice usable only by the person who uploaded it, and says the feature is for your use only. It is the honest route if the AI artist is really you behind a different name: your timbre, generated arrangements. It is not a way to borrow a real singer's voice, and you should not try.

Custom Models. Suno's release notes for v5.5 describe training a personalised model on at least six tracks from your own catalogue, and the help centre lists Custom Models for Pro and Premier subscribers. This locks the sound more than the voice: the arrangement habits and genre mix of your existing music. It suits someone who already has a body of work. An act that does not exist yet has nothing to train it on.

Which to use:

• No catalogue, no own voice in play: generate until one take has the voice, make a private Persona from it, keep the style string fixed.
• You sing, or want your own timbre: Voices.
• You already released six or more tracks in one consistent sound: a Custom Model for the sound, plus a Persona or Voices for the vocal.

None of these is a hard lock. A Persona leans the generation toward a voice; it does not freeze it. If you attach a soft alto Persona and then ask for aggressive metalcore, screamed vocals, you are handing the singer a song they cannot sing, and the result fights itself. The deeper mechanics of Personas and own-voice uploads are in the guide to Personas and your own voice in Suno, and the prompt side of keeping one singer is in the vocal persona lock guide.

The cheap backup for all three is a vocal anchor written into the style string itself: the same vocal description, word for word, at the front of every prompt. breathy female alto, close-mic, restrained in one session and soft girl vocals in the next is how a Persona gets overruled by your own wording.

Step 3: freeze the core sound, and decide in advance what may move

Most "my artist changed genre" problems are self-inflicted. A writer adds one descriptive word per session because the new song "needs a bit more energy", and after six releases the core style string has quietly grown four new tags. Here is the discipline that stops it.

Keep the core string short. A style prompt carrying fifteen descriptors gives the model fifteen ways to wander, and you will not notice which one moved. Six or seven strong tags that you never edit are steadier than a paragraph you rewrite each time.

Write the core in positive terms and put every "no" into the exclude field. "No rap, no autotune" inside the style prompt still puts the words rap and autotune in front of the model. If you want help building the string across every axis (genre, era, vocal, instruments, structure), the prompt generator at /generators does it deterministically, with the don'ts moved into Exclude. If you want a starting sound, the artist recipes describe more than 4 000 artists' sounds without using their names. That matters here for two reasons: Suno's help centre lists the names of well-known artists among the things that can block a generation, and an AI act that is openly "a copy of X" is exactly what streaming platforms and listeners push back on. Take a recipe as a starting point, then move two or three tags until the act sounds like itself.

Step 4: lock the words, which is the layer most AI artists forget

A consistent voice singing inconsistent lyrics still sounds like a playlist. Listeners hear an artist through what they write about and how: who is speaking, to whom, in which mood, with which objects. That is a brief, and a brief can be fixed.

This is where sunomarket fits, and where it does not. The song generator at /generators takes a form and returns a title and a full lyric with [Verse] and [Chorus] section tags, ready to paste into Suno's Lyrics field. It does not sing. There is no audio from us, ever: the voice, the Persona and the track all happen in Suno, which is a separate company we have no connection with and do not speak for. sunomarket is an independent site.

What the form gives an artist project is repeatability. For Lowtide Mira, the fixed fields would be:

POV        first
language   en
structure  ballad            (or classic, per the persona sheet)
hook       confession        (or title-line)
vocal      female
emotion    longing           (rotate: longing, loneliness, calm)
age        25-34, 35-44
theme      the only field that changes every song, plus the image bank

The theme is free text, so that is where each song gets its story and where you paste two or three objects from the image bank. Everything else stays put, which is what makes song six sound like the same writer as song one. A lyric that comes back looks like this (written for this article, not from any real song):

[Verse 1]
The bus to the coast runs twice on a Sunday
I know the times by the salt on the glass
You left me your coat and a map with the tide in it
I keep it folded like something that will pass

[Chorus]
I said I was leaving, I said it in June
The harbour still holds every word I don't mean

Run the result past your banned-words line before you paste it into Suno. A generator does not know your act's private rules unless you put them in the theme, so add "avoid the words fire, soul, forever" there, and check anyway.

Cost, so you can plan a release: SM1 is 20 credits per lyric, SM2 and SM3 are 80 credits and run an extra pass that flags weak lines and rewrites them. At 100 credits to the dollar that is $0.20 or $0.80 per lyric. If you draft five SM1 versions per single and keep one, that single's words cost 100 credits, a dollar. The $5 monthly plan gives 500 credits, which is 25 SM1 lyrics, and unspent plan credits clear at renewal; the $19 lifetime option is 1 500 credits once, 75 SM1 lyrics, that never expire. It is not unlimited, and new accounts start at zero credits, so there is no free trial of the lyric generator. Details are on /pricing.

If you prefer writing every word yourself, keep the form anyway as a checklist. The value is the fixed brief, not who types the lines.

Step 5: the name, the face, and the bio

The name is the one part of an AI artist you cannot fix later without losing whatever audience you built. Three checks before you commit: search the name on the streaming services you plan to release to, so your songs do not land on someone else's profile; check that the handle is free on the two or three social platforms you will actually use; and say it out loud in both languages you care about, so it does not turn into a joke in one of them. Generating and filtering an artist name covers the full set of checks and a brief that produces better candidates.

Never build the name, the avatar or the bio around a real person. An act called something close to a famous singer, with an avatar that looks like them, is exactly what Spotify's impersonation rule, described below, is aimed at, whatever the music sounds like.

Keep one avatar and one two-sentence bio everywhere. A bio that says plainly what the project is ("an AI-assisted project by a producer in Lisbon", or whatever is true for you) survives every policy change listed in the next section. A bio that implies a human singer who does not exist is the kind of thing that ages badly.

Step 6: release cadence and the disclosure rules that actually exist

Platform rules here are new and still moving, so here is what the platforms themselves say, at the time of writing, with no more than that.

• Spotify. Its September 2025 announcement says vocal impersonation is only allowed when the impersonated artist has authorised it, describes a spam filter aimed at mass uploads, duplicates and artificially short tracks, and backs an industry standard for AI disclosures in credits, built through DDEX and submitted by labels and distributors. Spotify says that disclosure is not about punishing artists who use AI responsibly.
• Deezer. Deezer labels albums that include fully AI-generated tracks and says it excludes those tracks from algorithmic and editorial recommendations. If your act is fully generated, plan on discovery coming from somewhere other than Deezer's recommendations.
• YouTube. The help page on altered or synthetic content names AI-generated music among the content creators have to disclose. You mark it during upload in YouTube Studio, with the AI-use setting in the Attributes section.
• Suno. The pricing page lists commercial use rights on the paid plans and "no commercial rights" on the free plan, and points to the terms of service for the limits. If the act is meant to earn, read those terms before release one, not after release ten.

Two practical rules follow from Spotify's spam filter alone. Do not release in bulk: an act that drops forty near-identical tracks in a month looks like the thing the filter was built to catch. And do not pad releases with very short tracks. A steady cadence (the sheet above says one single every four weeks) builds a catalogue that reads as an artist, not a feed.

Which distributors accept AI-assisted music, and which have removed it, is a question of its own: see which music distributor accepts AI songs. On ownership, your rights depend on the generator's terms and on the law where you live, and they differ between countries. This is not legal advice; our own position is on /disclaimer, and who owns AI-generated music walks through the licence-versus-copyright difference. Lyrics from our generator are non-exclusive and can overlap between users, which is one more reason to keep a banned-words line and an image bank of your own.

Where AI artists break, and the fix for each

The voice drifts on release three. Usually the style string changed, not the Persona. Diff the prompt you used against the persona sheet, word by word. If they match and it still drifts, the Persona was built from a weak take: generate a better anchor song and make "Mira v2", and note the version in the sheet.

Every song sounds the same. The lock is too wide. You froze the texture and the structure along with the genre. Move one texture and one structure choice into MAY VARY and rotate them on purpose.

Every song sounds different. The opposite: you are writing a fresh style string each session. Paste the core from the sheet, never retype it.

The lyrics are interchangeable. The brief has no image bank, or the theme field says "a sad song about the sea". Give the theme a specific situation and two objects, and keep the emotion inside the act's range.

The act looks like a copy of someone. You started from a famous reference and never moved off it. Change two production tags and one vocal descriptor until a listener could not name the source in one guess.

Your first week, if you start today

1. Write the persona sheet from Step 1. One page, all fields filled, including NEVER.
2. Build the core style string, with the exclude list, and test it on three throwaway lyrics. Keep generating until one take has the voice.
3. Make a private Persona from that take (or set up Voices, if the voice is yours). Write its name and version into the sheet.
4. Fix the lyric brief and write or generate the first single's words through it on /generators. Check them against the banned words.
5. Generate the single in Suno with the Persona and the untouched core string. Generate the second single the same day, so you can hear whether the act holds across two songs before anyone else does.
6. Pick and check the name, set one avatar and one honest bio, and choose a distributor.

If the second song sounds like the same artist as the first, you have an act. If it does not, fix the sheet, not the song.