Dev.to AI 🤖 Ai 👁 0 📖 4 min read

Stop Paying for Audio Stem Separation: Meet Split My Music (Self-Hosted, Local AI)

Have you ever tried to isolate vocals, drums, or bass from a song? If you have, you know the drill: you search online, land on a flashy web app, upload your audio, and immediately hit a wall. Either you get 2 free conve

Stop Paying for Audio Stem Separation: Meet Split My Music (Self-Hosted, Local AI)

Have you ever tried to isolate vocals, drums, or bass from a song?

If you have, you know the drill: you search online, land on a flashy web app, upload your audio, and immediately hit a wall. Either you get 2 free conversions before a paywall pops up, your file is compressed into a muddy 128kbps MP3, or you have to hand over your personal audio tracks to third-party cloud servers with questionable privacy practices.

A while ago, I built a tool called Voice Separator based on Meta’s Demucs. It did the job, but as Demucs began seeing less maintenance and my custom backend code started showing its age with tricky edge cases, I knew it was time for a fresh start.

I wanted something better. Something modern, privacy-first, lightning-fast, and dead-simple to self-host on a NAS, a homelab server, or a laptop.

Meet Split My Music.

👉 GitHub Repository: https://github.com/paladini/split-my-music

🐳 Docker Hub: paladini/split-my-music

What is Split My Music?

Split My Music is a lightweight, beautiful, self-hosted web application that lets anyone split audio files into vocals, instrumental music, and isolated instruments (drums, bass, guitars, piano) using state-of-the-art AI models running 100% locally.

No telemetry. No subscriptions. No cloud uploads. No arbitrary file size limits.

Split My Music Interface

Core Principles

  1. Simplicity Over Clutter: You shouldn’t need a degree in digital signal processing to isolate a vocal track. No confusing machine learning jargon in the primary interface — just clean labels like Vocals, Music, Drums, and Bass.
  2. 100% Local Processing: Every computation happens on your machine. Your audio never leaves your network.
  3. Hardware Agnostic: Runs smoothly on ordinary CPUs out of the box, with dedicated NVIDIA GPU acceleration (CUDA) when you want extra speed.
  4. Built on Giants: Instead of reinventing separation wheels, it is powered by the battle-tested python-audio-separator engine, supporting leading models like UVR-MDX-Net and Demucs v4.

Key Features

  • Drag & Drop Studio UI: Modern, responsive interface with drag-and-drop audio ingestion and dark/light ambient contrast.
  • Goal-Driven Model Selection: Choose between Vocals + Music (2 stems), Vocals + Drums + Bass + Other (4 stems), or full 6-stem isolation. Advanced users can still cherry-pick specific ONNX/PyTorch models.
  • In-Browser Audio Player & Waveform Cards: Listen to each separated channel side-by-side right inside the browser before downloading.
  • Multiple Export Formats: Export losslessly to 24-bit/32-bit WAV or FLAC, or high-bitrate MP3. Download individual channels or grab everything in a clean ZIP bundle.
  • Homelab & NAS Ready: Full Docker & Docker Compose support with persistent model caching and automatic cleanup of temporary files.

In-Browser Results

Up and Running in 30 Seconds

Getting started with Docker takes a single command.

Running on CPU:

docker run -d \
  --name split-my-music \
  -p 8000:8000 \
  -v split_models:/app/models \
  -v split_storage:/app/storage \
  paladini/split-my-music:latest

Or using Docker Compose:

services:
  split-my-music:
    image: paladini/split-my-music:latest
    container_name: split-my-music
    restart: unless-stopped
    ports:
      - "8000:8000"
    volumes:
      - models-cache:/app/models
      - storage-cache:/app/storage

volumes:
  models-cache:
  storage-cache:

Open your browser at http://localhost:8000, drop your favorite song, and watch the magic happen.

(Have an NVIDIA GPU? An override file docker-compose.gpu.yml is included in the repository with full CUDA 12.6 pass-through).

How It Was Built: Spec-Driven AI Engineering

Beyond the tool itself, the development journey was an experiment in modern software engineering.

I orchestrated the development using Google’s Antigravity environment paired with the TLC Spec-Driven methodology (by Felipe Rodrigues & the Tech Lead’s Club community).

Instead of prompting an AI with vague instructions and hoping for clean code, the spec-driven workflow enforced strict phases:

  1. Specify: Defining functional requirements (SEP-01 to SEP-12), edge cases, audio downmixing rules, and non-functional bounds.
  2. Design: Decoupling the engine adapter from FastAPI routes and creating a background job state machine with crash recovery.
  3. Tasks: Breaking implementation into small, atomic tasks with automated verification tests.
  4. Execute & Validate: Producing unit tests, real model integration tests, automated UI browser snapshots, and production-grade Docker pipelines.

In just a few hours, the process went from an empty directory to a complete application with automated CI/CD, documentation in English & Portuguese, test coverage, and pre-built multi-arch images on Docker Hub.

Why Open-Source Self-Hosted Software Matters

We are in an era where basic utilities are routinely turned into high-friction subscription services. Audio separation doesn’t need to be gated behind a monthly fee or contingent on uploading your files to third-party servers.

If you have a homelab, a NAS (like TrueNAS or Synology), or just a laptop, you have the right to own your tools and your data.

Give it a spin, star the project on GitHub, and let me know what you think! PRs, issues, and feature requests are very welcome.

👉 Project Link: https://github.com/paladini/split-my-music

👉 Docker Hub: hub.docker.com/r/paladini/split-my-music

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.