I built a music studio out of free software

What happened and why it matters

You do not need a studio, a record label, or a monthly subscription to make music any more. The whole studio runs on a graphics card you probably already own, it costs nothing, and once it is installed it never touches the internet again.

I want to be precise about what that sentence means, because it is easy to read it as advertising.

It means the thing that writes original music runs on your desk. The thing that splits a finished track back into separate voices, drums and bass runs on your desk. The final polishing step, the one that makes a track sit at the same volume as commercial releases, runs on your desk.

No account. No uploading. No terms of service that quietly claim a share of what you made.

Why that matters more than it sounds

For most of recorded history, that chain was the gatekeeper. Not talent. Not ideas. Access to a room full of equipment nobody could afford alone.

Then it became software that cost more than the computer. Then it became a subscription, which is the same gate with a friendlier face and a cancel button that turns your back catalogue into a hostage.

The part that changed my mind

The music program I used is free and open. The clever part is not simply that it makes music. It is that it comes with a planner.

You give it a feeling in plain words, something like "a slow rainy synthwave instrumental with a heavy low end", and a language model inside writes the actual detailed description, the tempo, the key and the length for you.

Then the chain continues, and it stays yours the whole way. One free tool pulls a finished track apart into its separate instruments, so you can mute the voice, keep the drums, or rebuild the bass. Another takes a track you love as a reference and matches yours to its volume and tone, which is the unglamorous step that separates something sounding like a demo from something sounding finished.

The honest catches

A guide that only tells you the good part is an advertisement, so here are four.

Quality varies, and it varies a lot. Some results are genuinely good. Some are mush. You are going to make several and keep one, which is exactly how studios have always worked, but it is worth knowing before you expect something wonderful on the first try.

The planner is only as good as the brain you give it. I ran that step through a small local model at first and the descriptions came out flat and generic, which produced flat and generic music. A stronger model wrote noticeably better descriptions and the output improved immediately. The bottleneck was never the music software. It was the thing writing its instructions.

The polishing step is not a mix engineer. It will make your track sit at commercial volume next to one you choose. It will not fix a bad arrangement, and it cannot tell you the tune is boring. Treat it as the last five percent, not the craft.

The lesson is the one this whole site keeps arriving at from different directions. The tools to make original work, and to actually own it, are free now and they run on hardware you already have.

How to do it yourself

Every tool here is free and open, every link goes to the official project, and nothing in this chain sends your work anywhere. Install from these places and nowhere else.

1. Check the hardware you already have

nvidia-smi

2. Install the room everything runs in

git clone https://github.com/Comfy-Org/ComfyUI
cd ComfyUI
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
python main.py

Official source: github.com/Comfy-Org/ComfyUI

3. Add the music model, and let it write the description for you

cd ComfyUI/custom_nodes
git clone https://github.com/ace-step/ACE-Step-ComfyUI
cd ACE-Step-ComfyUI
pip install -r requirements.txt

Now the part worth understanding. In the settings there is a switch that turns on the planner. Turn it on and you stop writing detailed descriptions altogether. You give it one line about the feeling you want, and it writes the rest itself. Leave a field blank and it fills that in too.

That single switch is the difference between operating a tool and describing an idea.

Official sources: the model and the pieces that run it

4. Split any track into its separate instruments

pip install -U demucs

demucs -n htdemucs mytrack.wav

# or just separate the voice, which is faster
demucs --two-stems=vocals mytrack.wav

Official source: github.com/facebookresearch/demucs

5. Polish it against a reference

pip install matchering

import matchering as mg
mg.process(
    target="mytrack.wav",
    reference="reference_i_like.wav",
    results=[mg.pcm24("mastered.wav")],
)

Official source: github.com/sergree/matchering

6. Put a local model in charge

curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen2.5:14b

Official sources: ollama.com and github.com/ollama/ollama

# 1. ask the local model to turn "rainy synthwave, heavy low end"
#    into a full description, plus tempo, key and length
# 2. put those values into the workflow
# 3. send it to the host program on your own machine
POST http://127.0.0.1:8188/prompt

# 4. check back until it has finished
GET  http://127.0.0.1:8188/history/<prompt_id>

# 5. read the filename from the result and hand over the file

7. Optionally teach it your own sound

The project documents how to train it further on a handful of your own songs, so it learns your style. Start with the official documentation rather than any third party guide, for the reason in the safety warning above.

8. Know where it lands

The end of all this is an ordinary audio file in a folder on your own disk. No export credits, no watermark, no account that can be closed, and no catalogue that disappears when a card expires.

You own the machine, the model, and the music.

Related: building your own AI agent is the same conductor idea applied to words, running your own assistant is where the local model habit starts, and the free software page tracks the open tools worth running yourself.