Proof of Use. An independent field guide · No. 01
Back to the collection
Music & Audio

Demucs

Separate a mixed music recording into vocals, drums, bass and other accompaniment with a locally runnable source-separation model.

stem separationvocalsmusic editinglocallimited maintenance
Source-reviewedNot hands-on testedSources checked

Start with an idea

What you can make

01

Isolate parts from your own recordings for practice and analysis

02

Prepare vocal and instrumental stems for permitted remix or editing work

Creator demonstration

See the documented workflow

Paper authors’ comparison samples, linked from the repository. Includes models beyond the standard downloadable package; no separation was run for this directory.

Compare the creators’ separated stems

Before you begin

What it takes to run

Python + audio dependencies

  • Current official continuation requires Python 3.10+ and PyTorch 2.1+
  • Use the documented uv/uvx or pip workflow; model checkpoints download separately
  • FFmpeg is recommended and needed for FLAC output or otherwise unsupported input formats
  • Intel Macs have an upstream Python/PyTorch version caveat; consult current instructions
Runs through / on

Linux · macOS · Windows · Command line · Python API

The real tradeoffs

Cost & licensing

Running costs

No code license fee for local separation; you supply the compute.

CPU mode is available. The README estimates roughly 7 GB GPU memory for default arguments, with lower-memory settings available; actual requirements and time depend on the model and track.

Code license

MIT

Read the code license

Models & other terms

Not separately specified in the inspected official HTDemucs weights repository

Read the model terms

Both code repositories contain MIT LICENSE files. The current official HTDemucs model card and file list omit a weights license; do not automatically extend the code license to every checkpoint. Verify weights terms before deployment. Input music rights are separate.

Know the limitations

  • Separation can leave bleed or artifacts; experimental piano separation is specifically caveated
  • Two-stem mode still performs full separation internally
  • Some sparse paper models are not supplied in the standard release
  • Archived original, limited-maintenance continuation and unclear standalone weights terms require extra diligence

Follow the evidence

Sources & verification

Official sources were reviewed on October 14, 2026. We did not install, run or benchmark this project. This is a source-check date, not a release date.

Links and project terms can change. These sources document the claims above; they are not a safety certification. Read our full method.

Keep exploring

A few other possibilities

All tools

05Music & Audio

Whisper

Transcribe speech in multiple languages or translate spoken audio into English with OpenAI’s locally runnable speech-recognition models.

Try it for

Create a reviewable transcript of your own interviews or podcasts

Read entry Official examples
Setup · Moderate
Python + FFmpeg
Cost to consider
Local compute

06Music & Audio

AudioCraft / MusicGen

Explore text- and melody-conditioned music generation with Meta’s AudioCraft research toolkit and MusicGen models.

Try it for

Sketch instrumental musical ideas for noncommercial experimentation

Read entry Official examples
Setup · Advanced
Python + GPU setup
Cost to consider
GPU compute; noncommercial weights

07Music & Audio

Chatterbox

Generate expressive speech from text with Resemble AI’s local text-to-speech family, including multilingual and reference-voice options.

Try it for

Prototype narration or spoken interfaces from your own scripts

Read entry Official examples
Setup · Moderate
Python + model download
Cost to consider
Local or rented compute