Make AI music on your own PC, songs with vocals included
Write a caption and some lyrics, and an open model sings them back to you as a finished song, on your own computer, with no credits to buy. ACE-Step 1.5 makes full songs with vocals under an MIT licence; MusicGen makes quick instrumentals. Learn to write captions that steer the style, split the result into stems, and know which licence lets you release what you make.
- Vocal and backing stems split with Demucs, ready for a remix or a karaoke version
- A clear licence table: what you can release, and what's for personal use only
- 16GB of RAM. ACE-Step runs on a CPU; an NVIDIA card with 6GB or more, or an Apple Silicon Mac, makes it far faster
- About 12GB of free disk: roughly 9.5GB of models plus Python packages
- uv from docs.astral.sh/uv ↗ and git: Install anything from GitHub
- Demucs for stems, set up in Remove the vocals from any song (optional)
Pick the model: ACE-Step, MusicGen or YuE
Three open music models, three different jobs and three different licences.
ACE-Step 1.5 (ACE Studio and StepFun) is the one to start with. A small language model plans the song (structure, lyrics timing, metadata), then a diffusion model renders the audio. It sings real lyrics in 50+ languages, makes 10 seconds to 10 minutes, and its model card ↗ says it runs in under 4GB of graphics memory. It's MIT licensed, and the card states that generated music can be used commercially.
MusicGen (Meta) makes instrumentals from a text prompt. It's simple and runs anywhere Python does, but its weights are CC-BY-NC 4.0: non-commercial use only. Great for learning and personal projects, not for anything you sell.
YuE2 writes a melody-and-chord plan and then a full song with vocals. It needs Linux, Python 3.12 and an NVIDIA card with 24GB, so it's a rented-GPU job (Rent a cloud GPU). Its weights are CC BY-NC 4.0 with an extra creator permission: individual creators may monetise outputs, companies need a licence from the authors (YuE on GitHub ↗).
Speed on an ordinary laptop, measured for this mission (Intel Core Ultra 7, 16GB RAM, no NVIDIA card): ACE-Step made 30 seconds of song in 640 seconds on the CPU; MusicGen small made 10 seconds in 159 seconds. On a graphics card both are many times faster: ACE-Step's authors quote under 10 seconds for a full song on an RTX 3090.
Get the kitmusic-kit.zip has the song script, the sample lyrics and the MusicGen script. kit.md has the caption cheat sheet, lyric tags and the licence table.