Demucs: Meta's AI-Powered Music Source Separation — Split Any Song into Stems

What Is Demucs?

Demucs is Meta's state-of-the-art music source separation tool. With over 10,000 GitHub stars, it uses a hybrid spectrogram and waveform deep learning model (Hybrid Transformer Demucs v4) to split any mixed audio track into individual stems — vocals, drums, bass, and other instruments — with remarkable quality.

Developed by Meta FAIR (Facebook AI Research), Demucs has evolved through four major versions. The latest Demucs v4 (HTDemucs) achieves the best separation quality in open-source, rivaling commercial solutions like Spleeter and iZotope RX. It's widely used by musicians, DJs, remixers, and content creators to create karaoke tracks, isolate vocals for sampling, and remix songs.

Why Demucs Dominates Music Separation

1. Hybrid Transformer Architecture (HTDemucs v4)

Demucs v4 uses a hybrid architecture that combines spectrogram-based frequency domain processing with waveform-based time domain processing. This dual approach captures both fine frequency details (important for separating overlapping instruments) and precise time structure (important for transient sounds like drums). The result is cleaner separation with fewer artifacts than pure spectrogram or pure waveform approaches.

2. Four-Stem Output

Demucs separates audio into four tracks: vocals, drums, bass, and other instruments. This 4-stem separation covers the most useful categories for remixing, karaoke creation, and music production. The "other" stem includes everything like guitars, keyboards, strings, and synths that don't fit into the first three categories.

3. GPU-Accelerated Processing

On a GPU, Demucs processes a 3-minute song in under 30 seconds — fast enough for interactive use. On CPU, it's slower (2-5 minutes per song) but still usable for batch processing. The v4 model is also more efficient than v3, requiring less memory while producing better quality.

4. Batch Processing

Demucs supports processing entire folders of audio files with a single command. Point it at a directory of MP3s and it'll separate every track, organizing each song into its own output folder with individual stem files.

5. MP4 and Video Support

Demucs can process video files directly, extracting the audio track, separating it, and optionally saving stems as separate audio files. This is perfect for YouTubers and video editors who need to isolate vocals from video clips.

Getting Started

Installation

pip install -U demucs
# Or clone for the latest:
git clone https://github.com/facebookresearch/demucs
cd demucs
pip install -e .

Basic Usage — CLI

# Separate a song into 4 stems
demucs song.mp3

# Output: separated/htdemucs/song/
#   - vocals.wav
#   - drums.wav
#   - bass.wav
#   - other.wav

Python API

import torch
import demucs.separator
from demucs import pretrained

# Load model
model = pretrained.get_model("htdemucs")
model.cuda()

# Separate audio
separator = demucs.separator.Separator(model)
sources = separator.separate_file("song.mp3")

# Access stems
vocals = sources["vocals"]    # shape: (channels, samples)
drums = sources["drums"]
bass = sources["bass"]
other = sources["other"]

# Save stems
from demucs.audio import AudioFile
AudioFile.save("vocals.wav", vocals.cpu(), model.samplerate)

Self-Hosting with Docker

docker pull joanfont/demucs:latest

docker run -d \
  --name demucs \
  --gpus all \
  -v ./input:/input \
  -v ./output:/output \
  joanfont/demucs:latest

Separate via Docker

docker exec demucs demucs /input/song.mp3 -o /output

Demucs vs Other Separators

Best Use Cases

Karaoke Creation

Remove vocals from songs to create instrumental karaoke tracks. Just separate and discard the vocals stem.

Sampling & Remixing

Isolate individual elements from existing songs — grab the drum track from one song, the bassline from another, and create something new.

Music Analysis

Analyze the structure of songs by examining separated stems. See how the drums, bass, and vocals interact throughout a track.

Audio Restoration

Separate audio to apply different processing to different elements — de-noise only the vocals while keeping the music untouched.

Tips for Best Results

• Use --two_stems=vocals to export only vocal and no-vocal stems (faster)
• Use the htdemucs model — it's the best quality model (v4)
• Use --mp3 flag to save stems as MP3 instead of WAV to save space
• Use --shifts=5 for higher quality (averages multiple predictions, ~2x slower)
• For GPU acceleration — ensure torch.cuda.is_available() returns True

Conclusion

Demucs is the gold standard for open-source music source separation. Meta FAIR's HTDemucs v4 delivers separation quality that rivals commercial tools, at no cost, with full GPU acceleration. Whether you're making karaoke tracks, pulling samples for hip-hop beats, or analyzing music structure, Demucs is the tool for the job.