How to make an AI song: a full guide to the AI song generator
An AI song generator is a program that writes the words, composes the music, performs the vocal and prints the cover of a song from a short description you type. Four years ago the first ones made thirty-second instrumentals. The 2026 generation (Suno, Udio, Mureka, Lyrumi) makes full songs with verses, choruses, bridges and a vocal that sings the lyrics you gave it. The difference between a toy and the kind of thing on a daily playlist is now about taste, not tech.
This guide walks through what one of them does when you press Make a song, what to type into the box so the result sounds like what you meant, and the handful of settings worth caring about. We build Lyrumi, so Lyrumi is the running example, but almost everything below applies to any of the modern AI song generators.
What an AI song generator actually does
Press the button on Lyrumi and four separate things happen, in order.
A language model writes the lyrics. It reads your line, picks a form (verse-chorus-verse, usually), decides how many beats a line should have, and writes the words. If you pasted your own lyrics it skips this step and keeps your words untouched.
A music model composes the arrangement. It chooses a key and a tempo that fit the mood of the lyrics, writes a melody for the vocal, and arranges the instruments around it.
A voice model performs the vocal. It reads the lyrics with the pitch and timing the composer wrote, in the voice you picked.
An image model paints the cover. It reads the title and the first few lines and paints a square image for the song.
All four happen in parallel after the first step, which is why a song is ready in roughly as long as it takes to put the kettle on.
No part of this is music retrieval. Nothing gets sampled from someone else's record. The vocal is synthesised the way a text-to-speech voice is synthesised, and the arrangement is generated note by note. That matters for the ownership question at the end.
How Lyrumi makes a song in three steps
In the app the whole thing is three fields and a button.
1. Say what the song is about
One or two sentences, in your own words. Not a prompt in a technical sense, just the thing you would tell a friend if you were about to describe a song to them. Real examples that have made songs people kept:
- the night the power went out and we played scrabble by candlelight
- my cat finally forgiving me for the move
- driving home at 4am after a wedding, every red light hitting right
- my grandmother's kitchen in the summer we visited every year
Each one has a specific object in it: a candle, a cat, a 4am red light, a kitchen. The AI picks up on specifics and leaves them in the song. "A song about love" produces a song that is also about love and about forty other things, which is to say about nothing much.
If you have the lyrics already (a poem, a chorus stuck in your head, something you wrote at sixteen), paste them in instead. The AI will not rewrite them. It will set them.
2. Pick a style, voice and tempo (or don't)
Three optional fields. Style is a sentence in your own words. "Indie folk, acoustic guitar, soft duet." "90s shoegaze, loud reverb." "Trap, slow, minor key." The style field reads whatever you write, so it is not a drop-down. Voice is female, male, duet, or instrumental (no vocal); the instrumental option is more useful than it sounds, because a song with no words is often what a video needs. Tempo is slow, medium or fast; a mood, not BPM.
Leave all three blank and Lyrumi reads your lyrics and picks for you. On a first song, do this. The result gives you a baseline to compare against later when you decide a specific voice or style is what you actually want.
3. Press Make a song
Lyrumi returns one take: two to three minutes long, an MP3 at 192 kbps, with a cover image. The take is private until you publish it. You can keep it to yourself, download it, publish it on your profile, or throw it out and make another with the same lyrics and different style settings, which Lyrumi calls a version and which costs the same as a fresh song.
Prompt tips that make a real difference
Write the way you talk. The box on the home page says Describe your song in one line. The models are tuned on how people actually describe songs to each other, which looks much more like the summer my brother moved out than like song:summer,brother,moved_out,bittersweet. Writing in commas does not help the model and it costs you the specifics that would have.
Name one thing. A song needs a hook, and a hook needs an image. "Love" is not an image. "The chipped mug she left on the counter" is an image, and the AI will put it in the chorus. If you cannot think of one, think of what a stranger would see in a photograph of the moment: a plant on a windowsill, a particular shoe, a kind of light.
Pick the voice second, not first. The order of the three fields is deliberate. The lyrics decide what voice makes sense, not the other way around. A song about your father in his garage is almost always a male vocal, and a lullaby in your mother's language is almost always a female one. If you pick a voice first and then write the line, you end up pushing the words toward the voice, and the words are the part that matters.
Say the key words of your genre once. The style field reads your sentence and matches it against the model's genre vocabulary. "90s shoegaze" is one of those pattern words, and it unlocks the pedal-chain sound. "Dreampop with heavy reverb" does the same. "A soft dreamy sound" doesn't, because it describes a feeling and not a genre.
Fixing a bad take
The first take is sometimes not it. The fix is usually one of three things.
The vocal is wrong: the pitch is right but the voice is not. Switch voice (male to female, or try the duet) and make a version. Keep the same lyrics and style.
The arrangement is wrong: the vocal is right but the backing is too loud, too fast, or the wrong genre. Rewrite the style field with the specific thing you want ("acoustic, no drums"; "half tempo") and make a version.
The lyrics are wrong: the AI misread the line. Edit your original description instead of the lyrics directly, and make a fresh song rather than a version. Describing the song better is more likely to help than editing the lyrics it wrote, because the lyrics are a shadow of your description.
Each version is 10 credits (one song), the same as a fresh render. A render that fails on our side is refunded in full and does not count against you.
Where to use the song
Three places people actually put them.
Keep it to yourself: the song stays in your Library and nobody else sees it. This is the default.
Publish it: it appears on your profile, in Explore in the app, and if the chart counts it in the week ahead, on lyrumi.com under your stage name. Taking it down removes it from all three immediately.
Download it: an MP3 at 192 kbps, no watermark. Use it in a video, send it to someone, put it on a USB stick for a car ride.
Published songs can get plays and likes from strangers. The home-page chart is counted, not curated, so what is on it is what people actually listen to. There is no editor and no featured slot. The first song you publish is listed in the app under Explore / New the moment it is done.
The question people ask next
The commercial question (can I sell this?) has its own guide, because the answer is longer than one paragraph: Can you sell AI-generated music?. The short version: on Lyrumi you own the songs you make, and whether they can be monetised on Spotify or YouTube depends more on those platforms' current rules than on the AI itself.
For the lyrics side specifically, lyrics to song AI is about pasting your own words in rather than letting the AI write them. If you are picking a tool and want a comparison, there is Best AI song generators in 2026 compared.
Your first song on Lyrumi is free. Open the app on Android or iPhone, type one line, and see where it ends up.
