Type the words, and an AI sings your song

Music Studio began as the answer to a different question: what if the AI wrote notes instead of audio, so every suggestion landed in a piano roll you could edit? That studio is still here, still free, and still the heart of the app. But it had one honest gap, and it was the gap this whole category of product is named for: there was no path from a page of lyrics to a produced song with a voice singing them. Now there is — and it was built the way everything else here is built, with the editable version of the truth kept in charge.

The sheet is what it sings

Every way words enter a track lands in one place: the karaoke sheet. Type them, paste them, fetch them from the web, transcribe them from a clip with AI — or press Write the words with AI and an original lyric is drafted from your track's own brief and theme. Structure tags like [Chorus] are stripped from the sheet, because nobody sings a stage direction, but they are kept for the song model, which reads them as arrangement hints.

That single sheet is then exactly what the sung render performs. There is no second "master lyrics" box that quietly drifts out of sync with what the karaoke view scrolls; the words you timed are the words you hear. Change a line, render again, and the two can never disagree, because there is only one of them.

A style in words the studio already knows

The render's style prompt starts from the same brief the free studio reads — the sentence you typed when you made the track — and always closes with the track's own facts: its tempo, its key, its scale. You can overwrite it, of course. And a style you like is worth keeping: saved styles live in one per-user library that appears as chips under the new-track brief and beside the paid style fields, because both sides of the studio speak the same language. One click refills either.

Knobs, but only where the model has them

Most song generators put the same three sliders on every render and let you discover, one credit at a time, which of them do anything. This one follows a stricter rule: only what the model actually honours is on the screen.

The sung model's one conditioning channel is words — so Voice (any, female, male) travels as words in the style text, which is the way that model listens. The instrumental and remix models expose real diffusion guidance — so Weirdness (how far the render may drift from literal adherence) and Style influence (how hard your style tags steer) are live exactly there, and greyed out on the sung road with a note saying why. A slider that silently does nothing would cost you a render to find out; a disabled one tells you before you spend.

Untouched knobs are not even sent: the model's own defaults stay in charge unless you actually moved something.

One take, one visible price

A click renders one take, and the card says what it costs before you press it. The result arrives as a muted clip in your arrangement with its own inline player: your composed track does not suddenly play twice, and the paid render is never inaudible. Like it? Unmute it in the Mix tab and it becomes the song. Prefer the version you sequenced by hand? The render sits there as an alternative, and the WAV export mixes whatever you left audible.

Remix anything the track owns

The remix card re-renders any clip this track already has — an upload, a mic take, an audio bed, an earlier AI render — toward a new style, with the project's own genre framing what the source was. It lands muted beside the original at the same bar, so A/B is one mute toggle. This is also what finally gives uploading a song a reason beyond karaoke: bring a rough demo, and re-hear it as house, as a ballad, as whatever the style box says.

Cover art that knows its place

One square image per track, generated from the title, genre and brief, shown on the track card and beside the render controls. The prompt explicitly forbids lettering — image models letter badly, and a cover is where it shows. The document's pointer to the art is server-owned, like everything else the page must not be able to forge.

Publish a score, not just a file

Publishing, from the transport bar, renders your arrangement and puts a snapshot on the community shelf: the audio plus the editable score — patterns, arrangement, mixer, lyrics. Later edits change nothing publicly until you republish; unpublish takes it back any time; and your private clips never travel.

On the shelf, tracks play in a persistent player with a queue, shuffle and repeat, and listeners keep private playlists. Plays count anonymously — a play is a fact, where a like is a vote and needs an account. And the button that matters most is Remix: it hands the published score to the listener's own Music Studio as a brand-new project. They start from the same notes, not the same recordings — which is exactly what a cover of a song has always been.

What stays free

The entire pattern studio: the sequencer, the piano roll, sections and drops, the mixer, karaoke with tap timing and .lrc export, built-in vocal removal, and the deterministic WAV export with its light master chain. Credits are spent only on the generated side — the sung render, the remix, the cover, the audio beds — and each card says so on its face. The free side is not a trial of the paid side; it is the other half of the same instrument, and the original tour of it still describes every corner.

Frequently asked questions

Do I have to write lyrics before I can render a song?

You need words for a sung render, because the model sings your sheet rather than inventing verses you never approved. The AI lyric writer will draft them from your brief in one click if you want a starting point. Or tick Instrumental and render without words at a length you choose.

What happens to the render — does it replace my track?

No. It arrives as a muted clip beside everything you sequenced, with its own player. Your arrangement keeps playing exactly as before until you unmute the render in the Mix tab. Keeping both, and A/B-ing them with one toggle, is the intended workflow.

Why are Weirdness and Style influence disabled on sung renders?

Because the sung model has no such parameters, and pretending otherwise would sell you a slider that does nothing. The instrumental and remix models do expose that guidance, so the sliders are live there. The voice control works on the sung road because that model reads its conditioning from the style words — which is where the setting actually goes.

If someone remixes my published track, what do they get?

The score: patterns, arrangement, mixer settings and lyrics, as a new project of their own. They never receive your recordings — not your voice takes, not your uploads, not your paid renders. Unpublishing removes the track and its score from the shelf entirely.

Does any of this change what the free studio can do?

No — that is the point of the design. The free studio still composes, plays and exports without spending anything, and the same document feeds both sides, with the same provenance the rest of Grafy keeps for reproducible workflows. A track can start as a sung render and end hand-finished in the piano roll, or start as patterns and end with a voice on top.

Open Music Studio and press Write the words with AI — or walk the whole surface first in the Music Studio tutorial.