Vocal Separation in 2 Minutes — Running Demucs and UVR Karaoke Locally

Hands-on results from running both locally. macOS Apple Silicon (CPU), Python 3.11.

Demucs — Two Lines, Done

Install:

pip install demucs torchcodec

torchcodec is needed by recent torchaudio for saving. Without it you get ImportError.

Run:

demucs song.mp3 --two-stems vocals

--two-stems vocals outputs vocal/accompaniment only instead of 4 stems. More practical for karaoke.

First run auto-downloads htdemucs model (~80MB).

Output:

separated/htdemucs/song/
├── vocals.wav
└── no_vocals.wav

Measurements: 10s audio → ~10s processing (CPU, Apple Silicon). Model: ~80MB, seconds to download.

UVR Karaoke — Bigger Model, Different Quality

Install:

pip install "audio-separator[cpu]"

Run:

audio-separator song.mp3 \
  --model_filename mel_band_roformer_karaoke_aufr33_viperx_sdr_10.1956.ckpt \
  --output_dir ./output

Long command — full model filename required. First run downloads 913MB model (1-5 min).

Output:

output/
├── song_(Instrumental)_...flac
└── song_(Vocals)_...flac

Measurements: 10s audio → ~43s (14s model load + 13s inference + save). Model: 913MB.

Comparison Table

Verdict

Quick test → Demucs. Simple install, short command, small model.
Karaoke quality → UVR Karaoke. Bigger and slower, but backing vocals stay in the accompaniment.