---

name: spleeter description: "Route Spleeter source-separation CLI, Python API, training, and MUSDB evaluation workflows." metadata: disco-role: operating disable-model-invocation: true license: MIT

Spleeter repo skill

Use this skill when the task names Spleeter, spleeter separate, spleeter train, spleeter evaluate, Deezer source separation, pretrained 2/4/5-stem music demixing, Separator, AudioAdapter, Spleeter JSON configs, MUSDB evaluation, or Spleeter-specific TensorFlow/ffmpeg/model-cache failures.

Spleeter 2.4.2 is a Python package for audio/music source separation using TensorFlow. It offers pretrained vocal/accompaniment and multi-stem models, a Typer CLI, a Python Separator API, training utilities for custom source-separation models, and MUSDB-style evaluation with optional extras.

Start here

1. Confirm runtime basics with scripts/check_install.py or these minimal checks:

   python -m spleeter --version
   python -m spleeter --help
   ffmpeg -version
   ffprobe -version
   

2. Read installation and runtime when setup, Python/TensorFlow versions, optional dependencies, ffmpeg, Windows, Apple Silicon, or GPU expectations matter.

3. Read CLI reference before constructing command-line calls.

4. Read models and configuration for spleeter: descriptors, model cache/download behavior, and custom JSON config loading.

5. For failures that are not clearly owned by one workflow, start with root troubleshooting, then route to the nearest sub-skill troubleshooting page.

Workflow routes

Task family	Read
Split local audio into vocals/accompaniment or 4/5 stems; choose pretrained descriptor; build spleeter separate commands; use Separator/AudioAdapter; handle filename templates, codecs, MWF, and output validation	separation
Prepare custom training CSVs/configs; validate data layout and STFT dimensions; generate tiny training fixtures; run or adapt spleeter train; troubleshoot caches/checkpoints/TensorFlow training	training
Evaluate a Spleeter model on MUSDB-style test/<song>/ data; install spleeter[evaluation]; interpret SDR/SAR/SIR/ISR; generate a tiny evaluation fixture; troubleshoot exit code 10, missing sources, and metrics output	evaluation

Capability boundaries

• Use this skill for operating Spleeter as a package, not for general audio DSP, generic TensorFlow training, unrelated source-separation libraries, or repository release engineering.
• The verified baseline is a CPU TensorFlow runtime with system ffmpeg/ffprobe. GPU acceleration can improve speed but is optional and must be verified in the user's active TensorFlow environment before promising GPU execution.
• Pretrained separation and evaluation may download model artifacts on first use. Treat first-run network/cache policy as part of the task plan.
• Spleeter's -i/--inputs style is deprecated in this version. Pass audio files as positional arguments to spleeter separate.
• Evaluation dependencies are optional: base Spleeter can import and run separation/training help, but spleeter evaluate requires musdb and museval from spleeter[evaluation].

Common entry points

Surface	Use
python -m spleeter separate -p spleeter:2stems -o separated song.wav	Basic vocal/accompaniment split; see separation workflows.
from spleeter.separator import Separator	Python API for in-memory waveforms and file outputs; see separation API.
python -m spleeter train -d DATA_ROOT -p CONFIG.json	Custom model training; validate first with training scripts.
python -m spleeter evaluate --mus_dir MUSDB_ROOT -o OUTPUT -p spleeter:4stems	MUSDB-style metrics; see evaluation workflow.
spleeter:2stems, spleeter:4stems, spleeter:5stems	Embedded pretrained descriptors; see models and configuration.

Evidence and provenance

This skill was distilled from Spleeter 2.4.2 source, package metadata, public README/CHANGELOG/notebook examples, embedded resource configs, and native tests. See repo provenance for the source commit and evidence paths, and repo routing metadata for managed router placement.