Demucs

Demucs

Demucs is Meta AI Research's music source separation model. It can split any music track into four stems — vocals, drums, bass, and other instruments — with state-of-the-art quality. Hybrid Demucs v4 combines waveform and spectrogram modelling for best-in-class performance on the SDR benchmark.

Key Features

• Four-stem separation (vocals, drums, bass, other) and six-stem variant
• Hybrid Demucs v4 sets the state of the art on the MusDB18-HQ benchmark
• Fine-grained model variants: htdemucs (default), htdemucs_ft, mdx, mdx_extra
• CLI for one-command separation; Python API for programmatic use
• Karaoke mode (exclude vocals) and guitar/piano two-stem separation

Quick Start

pip install demucs
demucs --mp3 mysong.mp3  # outputs to separated/htdemucs/

from demucs.apply import apply_model
from demucs.pretrained import get_model
from demucs.audio import AudioFile, save_audio

model = get_model("htdemucs")
audio = AudioFile("song.wav").read(streams=0, samplerate=model.samplerate, channels=model.audio_channels)
sources = apply_model(model, audio[None])
# sources shape: [batch, stems, channels, time]

npx ai-supply add demucs-music-source-separation

Curated mirror of the open-source Demucs (MIT). Get it from the source.