ElevenLabs Music: How to Make a Song with Your Own Lyrics

To make a song with your own lyrics in ElevenLabs Music, you need a single generation: write how it should sound in the main field (genre, tempo, instruments, voice) and paste the words into the “Lyrics” field, split into labelled [Verse] and [Chorus] sections. The model sings the text word for word, in English, Russian and other languages, and the length can be anything from 30 seconds to 3 minutes. In the IIshki studio it runs in the browser. Below: a lyrics template, style descriptions for different genres, five steps and the mistakes that make the model mix up verse and chorus or not sing at all.

How ElevenLabs Music turns lyrics into a song

The model card has two text fields, each with its own job.

• Description (the main field, up to 2000 characters): how the song sounds. Genre, mood, tempo, instruments, vocal type.
• Lyrics: what gets sung. The words are not translated or rewritten; the model performs them as they are.

Inside the lyrics the model looks for structure. Stanzas are separated by a blank line, and the first line of a stanza can be a section header in square brackets or after hashes: [Verse], ## Chorus. English and Russian names are recognized: intro, verse, pre-chorus, chorus (hook), bridge, outro (coda). Up to 12 sections per track. Headers are not sung; they only mark up the song.

Three more settings: length (30 seconds, 1, 2 or 3 minutes), engine (Music v2 is newer and cleaner and is the default; Music v1 is the previous generation, if you want a different sound) and the “Instrumental only” toggle.

Which AI to use for songs and sound

An overview of all music tools is in the roundup on making music and songs with AI.

Lyrics template

Copy it and replace the lines with yours. The blank line between blocks is required: that is how the model knows where a verse ends.

[Intro]
The night is calling me again

[Verse]
I'm driving through the empty town
The neon signs come rushing down
The city whispers, hold on tight
Our whole life's in the windshield light

[Chorus]
Don't let me go, don't let me go
While all the late-night streetlights glow
Don't let me go, don't let me go
Just one more mile before the snow

[Verse]
You said that time would wash it out
But every turn still spins me round

[Chorus]
Don't let me go, don't let me go
While all the late-night streetlights glow

[Outro]
Just one more mile before the snow

Performer notes like *(F)* or *(M)* are stripped and not sung, but the model will not split a duet by them: set the voice type in the description.

How to write lyrics that sing well

The model sings what it is given, so half of a song's quality is decided before you press generate.

• One idea per song. Sum up the story in one line: “a night drive through the city after a breakup”. Verses tell the story, the chorus repeats the main idea.
• A short, fixed chorus. Two to four lines, the same words every time. The melody sticks, and you can hear right away whether the model delivered it well.
• Steady meter. Keep paired lines at roughly the same syllable count with a clear rhyme scheme. Uneven lines give a choppy melody.
• Read it aloud. If a line is hard to say in one breath, shorten it. The model often swallows dense consonant clusters and long words.
• No extra symbols. Brackets, emoji and footnotes inside lines are not needed: service notes are stripped, and anything else the model may try to sing.

How to make a song with AI: 5 steps

1

1. Open the model

Go to the ElevenLabs Music card. Sign in with Telegram, email, Google or Yandex ID.

2

2. Describe the sound

In the main field, write 1–3 sentences on how the song should sound. The order “genre → mood → tempo → instruments → voice” works well. Examples:

Pop rock, nostalgic, mid tempo around 110 BPM, electric guitar and live drums, warm male vocal, big chorus

Acoustic ballad, tender and sad, slow tempo, piano and cello, female voice with a slight rasp

Danceable electropop for a party, 124 BPM, bright synths, four-on-the-floor kick, ringing female vocal, catchy hook

3

3. Paste the lyrics

Mark up the sections as in the template. Keep lines short and even: a line that is hard to sing in one breath will come out rushed. Write the chorus out in full each time instead of “chorus ×2”.

4

4. Choose the length

Match it to the text: a verse with a chorus fits 30 seconds to 1 minute, two verses with two choruses fit 2 minutes, a full song with a bridge fits 3. Time is split between sections, so too much text in a short version sounds rushed.

5

5. Draft first, then the final

Run 30 seconds or a minute first to hear the melody and voice. If you like it, generate the full length with the same description. The track lands in “My creations” and, if your account is linked to Telegram, the bot sends it as a file.

What to do with the finished song

• A singing photo. Upload a portrait and the track to Kling Avatar: the character sings with matching lip movement. More in our guide to talking photos with AI.
• A music video. Generate scenes in a video model and add the track in editing; to animate a frame, see how to animate a photo with AI.
• A voice-over clip. Use the track as a background for a video with narration, paired with MiniMax or ElevenLabs voice-over.

Common mistakes

What next

If you need a set of sounds for a clip rather than a track, try ElevenLabs Sound Effects. For narration without singing, see ElevenLabs Voiceover and the tips in our ElevenLabs guide.

The price depends on length and is shown on the model card before you start. No subscription is needed: tokens are bought once, and if a generation fails because of the model, they are refunded automatically.