Text To Song
Turns your lyrics, written with [verse] and [chorus] markers, into a sung track shaped by genre, instrument, mood and voice tags, using the open ACE-Step v1 3.5B music model.
Overview
ACE-Step v1 3.5B is an open music generation model developed by ACE Studio and StepFun and released under Apache 2.0. It pairs diffusion-based generation with Sana's Deep Compression AutoEncoder (DCAE) and a lightweight linear transformer, and writes complete songs, vocals included, at a duration you set.
On this page you supply the words. Lyrics go in as plain text with section markers, and five tag lists set the sound. ACE-Step names ten languages as its best performers: English, Chinese, Russian, Spanish, Japanese, German, French, Portuguese, Italian and Korean.
What you can build
- Demo vocals, so a songwriter can hear a draft lyric sung before recording it.
- Short custom songs for videos, podcasts or in-app moments, cut to an exact number of seconds.
- Lyric-to-song features in songwriting apps, with the style picked from menus.
Inputs
- Lyrics: plain text divided into sections with markers such as [verse], [chorus], [bridge] and [outro].
- Style tags, all five required, each allowing several picks: genre (well over 100 entries, from pop and jazz to bossa nova and 8-bit), instrument (piano, 808 bass, erhu, ukulele and many more), mood (nostalgic or epic, for example), voice (male, female, soprano, tenor, child, rapping) and vocal timbre (husky or airy, among others).
- Duration from 1 to 240 seconds (60 by default) and a seed, 42 unless you change it.
- Advanced controls: inference steps (1 to 500, default 60), guidance scale (1 to 30, default 15), euler or heun scheduling, three guidance types (apg, cfg, cfg_star), granularity and guidance interval settings, and separate weights for tags and lyrics (both 0 by default).
Outputs
Each run returns one MP3 with the vocals and backing mixed into a single stereo track. The page pairs the prefilled city-night lyric with a 60-second sample at 48 kHz.
Recommended settings
The defaults (60 steps, guidance 15, euler, apg) match ACE-Step's official demo, which marks euler and apg as the recommended options.
Lock the seed while you are still editing words or tags. Output moves a lot between seeds, and a fixed seed shows whether a change came from your edit. Once the lyric is final, try a handful of seeds and keep the best take.
To weight tags and lyrics separately, switch the guidance type to cfg. The demo's help text says the tag weight only works with cfg and suggests starting at 5.0 for tags and 1.5 for lyrics.
Limitations
The ACE-Step model card lists these weak points:
- Quality depends on the language; the ten named above do best.
- Rare instruments may not render well.
- Results swing with the seed and the requested duration, which the authors call "gacha-style" variation.
- Some styles underperform, with Chinese rap given as an example.
- Sung vocals can sound coarse.
- Fine control over individual musical parameters is limited.
Safety & compliance
The model is licensed under Apache 2.0. Its card lists three out-of-scope uses: generating copyrighted content without permission, producing harmful or offensive material, and presenting AI-generated music as human-made. The authors also ask users to verify that results are original, disclose AI involvement and respect cultural elements and copyright.
Example prompts
Great starting points for text-to-song-ACE-Step-v1-3.5B.
API quick start
Run text-to-song-ACE-Step-v1-3.5B with a single API call.
{
"prompt": "[verse]\nBright lights flashing in the cit…",
"genre": "Pop",
"instrument": "guitar",
"mood": "happy"
}curl -X POST "https://api.wiro.ai/v1/Run/ace-step/text-to-song-ace-step-v1-3-5b" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "[verse]\nBright lights flashing in the cit…",
"genre": "Pop",
"instrument": "guitar",
"mood": "happy"
}
JSONOther models you might like
Explore similar music generation models.



