How to Turn Lyrics Into a Song With AI: From Lyric Sheet to Finished Track

To turn lyrics into a song with AI, you need three inputs and one habit. The inputs: your lyrics split into labeled sections like [Verse] and [Chorus], a style prompt that describes the sound (genre, instruments, mood, voice), and a few settings you can mostly leave on default. The habit: generate several takes, keep the best one, and redo any section that didn't work on its own, rather than regenerating the whole song. The AI writes the melody, sings your words, and plays the backing track in a single pass.

This guide walks through that workflow from a finished lyric sheet to an exported track, using Song Creator Pro (a Windows app that runs the open ACE-Step 1.5 and YuE2 models offline). The formatting and prompting advice carries over to cloud tools too, and there is a section on those near the end. If your problem is the words themselves (lines that come out rushed, mispronounced names, a chorus that never lands), the companion article on getting better results from an AI song generator with your own lyrics covers those fixes one at a time.

What You Need Before You Start

• Finished or near-finished lyrics. Rough is fine, but you should know which lines are the verses and which are the chorus.
• A sense of the sound. One genre and a voice type is enough to start ("acoustic folk, warm male vocal").
• A generator. For this walkthrough, Song Creator Pro on a Windows PC.

Step 1: Pick the Model That Fits Your Lyrics

Song Creator Pro ships two models, and both turn lyrics into full songs with vocals. A few things decide which one to start with:

1. Language. YuE2 sings in English, Mandarin, Japanese, and Korean. ACE-Step 1.5 supports 50+ languages, so lyrics in Spanish, French, Hindi and so on go to ACE-Step.
2. How you want to edit afterwards. The Revise tool, which regenerates one section of a finished song, runs on ACE-Step 1.5. YuE2 instead (with planning left on Full, the default) writes each song as an editable score of melody, chords and lyrics that you can change and render again.
3. Structure. YuE2 writes out a melody-and-chord plan of your song before it renders any audio, so it follows your section tags closely. Our YuE2 prompting guide explains why.

If none of that decides it, try both on the same lyrics. Generation is local and unlimited, so a comparison costs time, not credits.

Step 2: Format Your Lyrics With Section Tags

Models read lyrics as a sequence of labeled sections, not as one long poem. The official docs for both models use bracketed section tags, and the ACE-Step 1.5 tutorial calls structure tags "the most powerful tool" in the lyrics field.

1. Put each section tag on its own line in square brackets. Stick to the standard names: [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Outro].
2. Leave a blank line between sections. The ACE-Step tutorial gives this as its formatting rule, and our YuE2 prompting guide gives the same one.
3. Write every repeated chorus out in full. In our testing, and in what other users have reported, YuE2 does not expand shorthand like "(repeat chorus)" or "x2"; it tries to sing those words. Writing it out is the safe habit with either model.
4. Keep everything that isn't sung out of the field. No song title, no "(guitar solo here)", no notes to yourself. YuE2's official agent guide says to keep implementation notes out of the lyrics, and with ACE-Step, stacking too many tags can make the model treat tag text as words to sing.
5. Keep sections short. Four to six lines per section is a reliable range in practice. A verse of sixteen long lines is where vocals start to rush.

A lyric sheet ready to paste:

[Verse]
I left the porch light on for you
The way we always used to do
The kettle's cold, the radio's low
I'm still not sure you'll even show

[Chorus]
So come home slow, come home late
I don't mind how long I wait
The door's unlocked, the light is on
Come home slow, but come home

[Verse]
The calendar is two weeks wrong
I never turned it, it's been long
Your coat is hanging where it was
I leave it there, I guess because

[Chorus]
So come home slow, come home late
I don't mind how long I wait
The door's unlocked, the light is on
Come home slow, but come home

[Outro]
Come home slow, but come home

Notice what the lines have in common: they are short, similar in length, and the chorus is identical both times. That is not an accident, and the lyrics tips article explains each choice.

A few formatting extras are model-specific. With ACE-Step 1.5 you can add a short modifier to a tag ([Chorus - anthemic]), insert [Instrumental] between sections for a break, use UPPERCASE for more vocal intensity, and put words in parentheses to have them treated as background vocals. With YuE2, keep to plain section tags: anything in the lyrics field that isn't a tag gets planned as sung words, parentheses included.

Step 3: Describe the Sound in the Style Prompt

The Description field (the style prompt) controls the sound. The lyrics carry the story; the description tells the model what it should sound like. Cover four things:

1. Genre, ideally a specific one: "90s alt-country" beats "country".
2. Instruments: "fingerpicked acoustic guitar, brushed drums, pedal steel".
3. Mood: one or two words, not five.
4. Voice: gender and tone, such as "warm male vocal" or "breathy female vocal". Leave it out and the model picks the voice for you, and the voice is what listeners notice first.

For the lyrics above:

slow folk ballad, fingerpicked acoustic guitar, soft upright bass, brushed drums, warm intimate male vocal, nostalgic

Don't restate the story in the description ("a song about waiting for someone to come home"). That belongs in the lyrics, where it already is. With ACE-Step, the AI Enhance toggle reformats your description into the structure the model reads best. With YuE2, write comma-separated tags and put the most important ones first. YuE2's official docs place the language in the style prompt, so add it as a tag ("Japanese", "Mandarin"), especially when you are not singing in English.

For deeper dives, see the AI music prompting guide (ACE-Step) and the YuE2 prompting guide.

Step 4: Leave Most Settings on Default

Under Music Details you'll find BPM, Duration, Key, Time Signature, and Language. For a first pass, leave them blank and let the model infer them from your description and lyrics. Set one only when you need it:

1. Duration, when the song needs a specific length. Make sure your lyrics fit it: a long duration with two short verses leaves long instrumental stretches, and a short duration with many verses forces rushed vocals. As a rough guide, plan for 2 to 3 sung words per second, which is about 90 to 140 words for a 47-second track or 200 to 300 for a 2-minute song.
2. BPM, when you need a precise tempo. Keep it consistent with the description; 150 BPM on a "slow ballad" prompt sends mixed signals.
3. Language, if the vocals come out in the wrong language. ACE-Step usually detects the language from your lyrics, so this setting is mainly a fix for when it guesses wrong.

For ACE-Step, the official tutorial notes that short songs (30 to 60 seconds) and medium lengths (2 to 4 minutes) are the most stable, and that very long generations are more prone to repetition or structure problems.

Step 5: Generate a Batch, Not a Single Take

Open Advanced Settings and set Batch Size to 2 to 4, with Seed left at -1 (random). Then hit Generate.

Each take gets a different random seed, so the same lyrics come back with different melodies, phrasing, and vocal performances. The first take is rarely the best. Both models' developers work this way: the ACE-Step tutorial suggests larger batch sizes so you can compare versions, and YuE2's published benchmark scores come from picking the best of two or eight candidates.

Step 6: Listen for the Right Things

When the batch finishes, listen to each take with your lyric sheet open and check, in this order:

1. Are the words right? Every line is sung, nothing is skipped or invented, and names sound close enough to how you say them.
2. Is the structure right? The verses sound like verses, the chorus lifts, and the song ends where it should.
3. Does the chorus land? This is the part people remember. A take with a great chorus and a shaky verse is often worth keeping, because the verse can be fixed (Step 8).
4. Is it the right voice and sound? If every take misses on this, the description needs work, not more takes.

Pick the best take as your keeper and click Save to export it to your computer before you change anything.

Step 7: Iterate With Seeds and Batches

What you change next depends on what missed.

1. The sound is right but no take is good enough: generate another batch with the same inputs and the seed left on -1, so every take gets a new random seed.
2. The sound is wrong in every take: change the description. To compare fairly, set Seed to a fixed number so the only thing that changed between runs is your prompt. The ACE-Step tutorial recommends fixing the seed for exactly this reason. Set it back to -1 when you want variety again.
3. The same line comes out rushed or garbled in every take: the problem is the line itself, not bad luck. Shorten it or respell the problem word, then regenerate. The lyrics tips article covers both fixes.
4. You want to reproduce a take: reuse its seed with the same inputs. Results may still differ slightly on different hardware or settings, so save any take you like instead of counting on regenerating it.

Step 8: Fix One Section With Revise

This is the step that saves the most time. If the song is 90% right and one verse is weak, don't regenerate the whole thing and risk losing the parts that already work. Use Revise, which regenerates a selected stretch of the track while keeping everything around it.

1. Click Revise on your keeper (or open the Revise tab and upload the file). Revise runs on ACE-Step 1.5 and accepts any audio, including a song you made with YuE2.
2. Drag the two handles on the waveform to select the section to redo. Revise works on stretches between 3 and 90 seconds.
3. Paste the song's full lyrics into the Lyrics field with the same section tags, then edit only the lines in the section you are fixing. Edit words, not tags.
4. Leave Revise Mode on Balanced and Revise Strength around 0.4 to 0.6 for lyric changes.
5. Set batch size to 2 or 3 and generate. Keep the version that blends in most naturally.

If your rewritten lines are about the same length as the originals, the new section tends to blend in better. Our full guide to revising part of a song with AI covers the blending settings for when the edges sound abrupt.

On YuE2 songs you have a second option: open the song's score, change the lyrics there, and render it again. That rebuilds the audio from the edited score rather than patching one stretch of it.

Step 9: Export, or Take It Further

When you have a version you're happy with, Save it as an MP3, or as FLAC or WAV if you plan to edit or master it. From there:

• Record your own vocal over the AI backing by extracting the instrumental with the Extract tool.
• Add instruments around an existing track with Layer.
• Reimagine the whole song in a new genre with Remix.

Doing This in a Cloud Generator Instead

You don't need a local app to turn lyrics into a song. Suno accepts your own lyrics in its Custom mode, works in any browser including on a phone, and its help center says Pro and Premier subscribers can use Replace Section to change lyrics in one part of a song. If you only have a phone or a Mac, a cloud tool is the practical choice. The section-tag and line-length advice in this guide applies there too.

The trade-offs are download limits and the cost of each take. As of 16 October 2026, Suno's pricing page lists no monthly song downloads on the free plan, 20 a month on Pro, and 60 a month on Premier, and every generation spends credits. Udio switched downloads off when it settled with Universal Music Group on 29 October 2025, and they are still off. Getting a good sung take of your own lyrics usually takes several attempts, so paying per generation adds up. A local app like Song Creator Pro is a one-time purchase with no generation or download limits, but it is Windows only, works best with a GPU, and has no browser or mobile version. Our cloud vs local comparison goes deeper.

Who Owns a Song Made From Your Lyrics?

Writing the lyrics yourself matters here. The US Copyright Office's position is that purely AI-generated material without human authorship is not copyrightable, while human contributions such as lyrics you wrote can be protected. Song Creator Pro includes a commercial license for what you generate. For the full picture, including selling and distributing, see who owns AI-generated music.