How to make a song with AI with your own lyrics (in about two minutes, for about 9 cents)
Also in: Português
To make a song with AI from your own lyrics, open Music Studio on rfab.ai, pick "I write the lyrics", paste your lyrics, press "Split into sections", pick a model and describe the sound, then press "Generate song". On Mureka V9 one run costs 4,500 tokens (about $0.09), gives you two takes, and took about 90 seconds in our test. Your words are sung exactly as you wrote them.
1. Open Music Studio
Go to rfab.ai/music-studio and sign in. A new account gets 150,000 free tokens. There are five ways to start. To use your own words, pick I write the lyrics.
2. Paste your lyrics
Paste your lyrics in the box. Put a blank line between the parts of the song. You can label each part in square brackets: [Verse], [Pre-Chorus], [Chorus], [Bridge]. If you do not label them, a part that repeats word for word becomes the chorus.
3. Split into sections
Press Split into sections. It is free, and no AI touches your words. Each part becomes a section with a type and a label, and you can change a wrong type. The Rewrite, Tighten and Hooky buttons under each section are optional: they ask an AI to change that section, and they cost tokens.
4. Title, length and language
Click Sound. Give the song a title, pick a length (Jingle 0:30, Short 1:30, Standard 3:00 or Long 4:30) and set the language of your lyrics. Most models sing at their own pace, so the length is a target.
5. Pick a model
Every model sings your own lyrics.
| Model | Price | Note |
|---|---|---|
| Mureka V9 (ticked when the page opens) | 4,500 tokens (about $0.09) | Two takes per run |
| Lyria 3 Pro (Google) | 8,000 tokens | Up to 3:00 |
| MiniMax Music 2.6 | 15,000 tokens | Up to 5:00 |
| Eleven Music v2 | 250 tokens a second (3:00 = 45,000) | Follows the length exactly |
Choose two models to hear the same words sung two ways in one run.
6. Describe the sound
Type a genre and press Enter, then the mood and the instruments. Pick the voice and the tempo. Anything else goes in the last box. Describe the sound, not an artist: Lyria and Eleven Music refuse a real artist's name.
7. Generate, listen, download
The button shows the price before you pay. Press Generate song. Play both takes: the words light up as they are sung. A model can skip a line or sing a part twice, so keep the take that sings your words best. Download the mp3 or the wav, and the timed lyrics as LRC or SRT.
What it costs
50,000 tokens = $1. One run on Mureka V9 = 4,500 tokens for two takes. The 150,000 free tokens of a new account make about 33 runs, and you get more free tokens every day.
Common problems and fixes
- The Generate button is grey. The page says why: pick at least one model, add lyrics, or sign in.
- Lyria 3 (30 s clip) refuses the song. Set Length to Jingle (0:30).
- "looks like an artist name". Take the name out and describe the sound instead.
- A provider blocks a song. Those tokens are not refunded. Every other failure is refunded.
Tips for beginners
- Keep each line short, about 8 syllables. A long line gets rushed.
- Write the chorus the same way every time it comes back.
- Use the English labels (
[Verse],[Chorus]). The page reads them for lyrics in every language. - Start on Mureka V9: five tries cost less than one song on Eleven Music.
- Set the language, and the model sings in Portuguese, Spanish, French, German and more.
Advanced tips
- Stems (10,000 tokens) splits a Mureka song into vocals, drums, bass and the rest.
- Extend continues a Mureka song from the second you pick. Loop and Trim are free.
- Make the video turns the song into a simple video (one moving picture under the whole song) or a full music video with a storyboard.
- The API: send your sections and style to
POST /api/music/generate, or use the RFab command line tool:rfab music make --spec song.json --models mureka:mureka-9 --out song.
The song in the video, "Sing It Back to Me", was made this way on Mureka V9 in one run. The bonus at the end, "Canta pra Mim", is the same song in Portuguese, made with the same steps.