New AI Music Model Can Write 5-Minute Songs From a Text Prompt
Type a description, get a finished song up to five and a half minutes long.
StepAudio 3 Music is a new AI system that writes music from a text description. Ask it for a slow country ballad about a rainy town, and it produces a finished track. It handles full songs with singing, instrumentals, backing tracks built from a vocal recorded with no music behind it, and cover versions. The longest output is five minutes and 30 seconds, which is a real song rather than a loop.
The clever part is that it plans before it plays. One half of the system sketches the song in ABC notation (a simple text shorthand for notes), which keeps the harmony, rhythm and melody from drifting apart over several minutes. The other half turns that plan into actual sound at 48-kHz quality, which is sharper than a CD. The team then trained it on human preferences — the same trick used to make chatbots pleasant, where people pick which of two outputs is better.
How good is it? In the team's own tests, it scored highest on enjoyment, usefulness and production quality, and on a measure of how well the music matches its text prompt. On the independent Artificial Analysis Music Arena vocals leaderboard it earned a quality score of 1105 — third place, behind Suno V5.5 and Mureka, and ahead of Suno V5, MiniMax and others. So it is near the top of the pile, but not yet the champion.
For everyday listeners and creators, this means custom music keeps getting faster and cheaper. A podcast intro, a game soundtrack or a wedding slideshow tune no longer requires hiring anyone. But two caveats: this is a technical report, not a product you can download today, and AI covers and training data still sit in a legal grey zone. People who sell stock or session music may feel the squeeze first.
- It turns a plain-English description into a finished song, up to five and a half minutes long — no instruments or musicians needed.
- It ranked third on a public vocal-quality leaderboard (score 1105), behind Suno V5.5 and Mureka but ahead of Suno V5.
- It also makes backing tracks from a bare vocal and cover versions, which is handy for songwriters and uncomfortable for session musicians.
Why It Matters
Custom music could soon cost pennies and take seconds — great for creators, tough for people who sell stock tunes.